[llvm-commits] [llvm] r107613 - /llvm/trunk/lib/Target/X86/README-SSE.txt

[Eli Friedman]

On Sun, Jul 4, 2010 at 10:48 PM, Chris Lattner <sabre at nondot.org> wrote:
> Author: lattner
> Date: Mon Jul  5 00:48:41 2010
> New Revision: 107613
>
> URL: http://llvm.org/viewvc/llvm-project?rev=107613&view=rev
> Log:
> some notes about suboptimal insertps's
>
> Modified:
>    llvm/trunk/lib/Target/X86/README-SSE.txt
>
> Modified: llvm/trunk/lib/Target/X86/README-SSE.txt
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/README-SSE.txt?rev=107613&r1=107612&r2=107613&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Target/X86/README-SSE.txt (original)
> +++ llvm/trunk/lib/Target/X86/README-SSE.txt Mon Jul  5 00:48:41 2010
> @@ -846,3 +846,34 @@
>  doing a shuffle from v[1] to v[0] then a float store.
>
>  //===---------------------------------------------------------------------===//
> +
> +On SSE4 machines, we compile this code:
> +
> +define <2 x float> @test2(<2 x float> %Q, <2 x float> %R,
> +       <2 x float> *%P) nounwind {
> +  %Z = fadd <2 x float> %Q, %R
> +
> +  store <2 x float> %Z, <2 x float> *%P
> +  ret <2 x float> %Z
> +}
> +
> +into:
> +
> +_test2:                                 ## @test2
> +## BB#0:
> +       insertps        $0, %xmm2, %xmm2
> +       insertps        $16, %xmm3, %xmm2
> +       insertps        $0, %xmm0, %xmm3
> +       insertps        $16, %xmm1, %xmm3
> +       addps   %xmm2, %xmm3
> +       movq    %xmm3, (%rdi)
> +       movaps  %xmm3, %xmm0
> +       pshufd  $1, %xmm3, %xmm1
> +                                        ## kill: XMM1<def> XMM1<kill>
> +       ret
> +
> +The insertps's of $0 are pointless complex copies.

I'm also concerned about the pshufd at the end; is the ABI stuff
really working the way you want it to?

-Eli