[Mesa-dev] [PATCH] R600: Use MULADD_IEEE instruction for mad pattern

[Michel Dänzer]

On Don, 2013-02-07 at 20:17 +0100, Vincent Lejeune wrote: 
> 
> diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
> index 85bb73a..65788f8 100644
> --- a/lib/Target/R600/SIInstructions.td
> +++ b/lib/Target/R600/SIInstructions.td
> @@ -1360,7 +1360,7 @@ def : Pat <
>  /**********   VOP3 Patterns    **********/
>  /********** ================== **********/
>  
> -def : Pat <(f32 (IL_mad AllReg_32:$src0, VReg_32:$src1, VReg_32:$src2)),
> +def : Pat <(f32 (fadd (fmul AllReg_32:$src0, VReg_32:$src1), VReg_32:$src2)),
>             (V_MAD_LEGACY_F32 AllReg_32:$src0, VReg_32:$src1, VReg_32:$src2,
>              0, 0, 0, 0)>;

This breaks 36 piglit tests with radeonsi, just like my patch
http://lists.freedesktop.org/archives/mesa-dev/2013-January/033066.html
did. The only difference in the generated code is that your patch uses
V_MAD_LEGACY_F32 (which I don't think makes sense for (fadd (fmul)))
while mine used V_MAD_F32. Same problem using V_FMA_F32 as well. Still
not sure why it breaks. :(


-- 
Earthling Michel Dänzer           |                   http://www.amd.com
Libre software enthusiast         |          Debian, X and DRI developer