[Mesa-dev] [PATCH] gallium/radeon: fix behavior of GLSL findLSB(0)

[Nicolai Hähnle]

On 26.10.2016 01:17, Marek Olšák wrote:
> From: Marek Olšák <marek.olsak at amd.com>
>
> This is for 12.0 and older. A different commit fixes 13.0 and newer.
>
> Cc: 11.2 12.0 <mesa-stable at lists.freedesktop.org>

Reviewed-by: Nicolai Hähnle <nicolai.haehnle at amd.com>

> ---
>  src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c | 19 ++++++++++++++-----
>  1 file changed, 14 insertions(+), 5 deletions(-)
>
> diff --git a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
> index 74b36ec..9893bfe 100644
> --- a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
> +++ b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
> @@ -1326,23 +1326,32 @@ static void emit_lsb(const struct lp_build_tgsi_action * action,
>  		     struct lp_build_emit_data * emit_data)
>  {
>  	struct gallivm_state *gallivm = bld_base->base.gallivm;
> +	LLVMBuilderRef builder = gallivm->builder;
>  	LLVMValueRef args[2] = {
>  		emit_data->args[0],
>
>  		/* The value of 1 means that ffs(x=0) = undef, so LLVM won't
>  		 * add special code to check for x=0. The reason is that
>  		 * the LLVM behavior for x=0 is different from what we
> -		 * need here.
> -		 *
> -		 * The hardware already implements the correct behavior.
> +		 * need here. However, LLVM also assumes that ffs(x) is
> +		 * in [0, 31], but GLSL expects that ffs(0) = -1, so
> +		 * a conditional assignment to handle 0 is still required.
>  		 */
> -		lp_build_const_int32(gallivm, 1)
> +		LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
>  	};
>
> -	emit_data->output[emit_data->chan] =
> +	LLVMValueRef lsb =
>  		lp_build_intrinsic(gallivm->builder, "llvm.cttz.i32",
>  				emit_data->dst_type, args, ARRAY_SIZE(args),
>  				LLVMReadNoneAttribute);
> +
> +	/* TODO: We need an intrinsic to skip this conditional. */
> +	/* Check for zero: */
> +	emit_data->output[emit_data->chan] =
> +		LLVMBuildSelect(builder,
> +				LLVMBuildICmp(builder, LLVMIntEQ, args[0],
> +					      bld_base->uint_bld.zero, ""),
> +				lp_build_const_int32(gallivm, -1), lsb, "");
>  }
>
>  /* Find the last bit set. */
>