[llvm] [LoongArch] Optimize extractelement containing variable index (PR #151475)

[via llvm-commits]

tangaac wrote:

> > Are these okay?
> > ```
> > xr0: v32i8
> > a0 : index
> >                                    L/T
> > xvpermi.q       xr1, xr0, 1        3/4
> > xvreplgr2vr.b   xr2, a0            3/1  // movgr2fr.w    f2, a0    2/1
> > xvshuf.b        xr2, xr1, xr0, xr2 1/2
> > Total:                             7/1
> > 
> > xr0: v16i16
> > a0 : index
> >                                    L/T
> > xvpermi.q       xr1, xr0, 1        3/4
> > xvreplgr2vr.h   xr2, a0            3/1  // movgr2fr.w    f2, a0    2/1
> > xvshuf.h        xr2, xr1, xr0      1/2
> > Total:                             7/1
> > 
> > xr0: v4i64
> > a0 : index
> >                                    L/T
> > xvpermi.q       xr1, xr0, 1        3/4
> > xvreplgr2vr.d   xr2, a0            2/1
> > xvshuf.d        xr2, xr1, xr0      1/2
> > Total:                             6/1
> > ```
> 
> I think it is right and better. My thought process was entirely limited by `xvperm.w`. Thanks for your idea.
> 
> What do you think about this? @tangaac

It shows better performance.

https://github.com/llvm/llvm-project/pull/151475