[Beignet] [PATCH] Refine the FCMP_ORD and FCMP_UNO.

[Zhigang Gong]

LGTM, pushed, thanks.

On Mon, Mar 24, 2014 at 05:21:40PM +0800, Yang Rong wrote:
> If there is a constant between src0 and src1 of FCMP_ORD/FCMP_UNO, the constant
> value must be ordered, otherwise, llvm will optimize the instruction to ture/false.
> So discard this constant value, only compare the other src.
> 
> Signed-off-by: Yang Rong <rong.r.yang at intel.com>
> ---
>  backend/src/llvm/llvm_gen_backend.cpp | 22 +++++++++++++++++++---
>  1 file changed, 19 insertions(+), 3 deletions(-)
> 
> diff --git a/backend/src/llvm/llvm_gen_backend.cpp b/backend/src/llvm/llvm_gen_backend.cpp
> index c459f25..2904037 100644
> --- a/backend/src/llvm/llvm_gen_backend.cpp
> +++ b/backend/src/llvm/llvm_gen_backend.cpp
> @@ -1653,10 +1653,26 @@ namespace gbe
>        case ICmpInst::FCMP_OGE: ctx.GE(type, dst, src0, src1); break;
>        case ICmpInst::FCMP_OLT: ctx.LT(type, dst, src0, src1); break;
>        case ICmpInst::FCMP_OGT: ctx.GT(type, dst, src0, src1); break;
> -      case ICmpInst::FCMP_ORD: ctx.ORD(type, dst, src0, src1); break;
> +      case ICmpInst::FCMP_ORD:
> +        //If there is a constant between src0 and src1, this constant value
> +        //must ordered, otherwise, llvm will optimize the instruction to ture.
> +        //So discard this constant value, only compare the other src.
> +        if(isa<ConstantFP>(I.getOperand(0)))
> +          ctx.EQ(type, dst, src1, src1);
> +        else if(isa<ConstantFP>(I.getOperand(1)))
> +          ctx.EQ(type, dst, src0, src0);
> +        else
> +          ctx.ORD(type, dst, src0, src1);
> +        break;
>        case ICmpInst::FCMP_UNO:
> -        ctx.ORD(type, tmp, src0, src1);
> -        ctx.XOR(insnType, dst, tmp, getRegister(cv));  //TODO: Use NOT directly
> +        if(isa<ConstantFP>(I.getOperand(0)))
> +          ctx.NE(type, dst, src1, src1);
> +        else if(isa<ConstantFP>(I.getOperand(1)))
> +          ctx.NE(type, dst, src0, src0);
> +        else {
> +          ctx.ORD(type, tmp, src0, src1);
> +          ctx.XOR(insnType, dst, tmp, getRegister(cv));  //TODO: Use NOT directly
> +        }
>          break;
>        case ICmpInst::FCMP_UEQ:
>          ctx.NE(type, tmp, src0, src1);
> -- 
> 1.8.3.2
> 
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet