[Mesa-dev] [PATCH 05/15] ac: revert new LLVM 7.0 behavior for fdiv

[Timothy Arceri]

Hi Marek,

It's late and I haven't dug into this any further but this patch causes 
a whole bunch of f64 piglit tests to fail for the radeonsi nir backend.

e.g.

./bin/shader_runner 
generated_tests/spec/glsl-4.00/execution/built-in-functions/fs-inverse-dmat2.shader_test 
-auto -fbo


LLVM ERROR: Cannot select: 0x7fbc48075aa8: f64 = bitcast 0x7fbc48077730
   0x7fbc48077730: f32 = RCP 0x7fbc4806e788
     0x7fbc4806e788: f64 = fadd nsz 0x7fbc480757d0, 0x7fbc48075a40
       0x7fbc480757d0: f64 = fmul nsz 0x7fbc4806f0e0, 0x7fbc4806f420
         0x7fbc4806f0e0: f64 = bitcast 0x7fbc4806f078

On 30/8/18 6:13 am, Marek Olšák wrote:
> From: Marek Olšák <marek.olsak at amd.com>
> 
> Cc: 18.1 18.2 <mesa-stable at lists.freedesktop.org>
> ---
>   src/amd/common/ac_llvm_build.c | 9 ++++++++-
>   1 file changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
> index c741a1ab62d..629cd2a7527 100644
> --- a/src/amd/common/ac_llvm_build.c
> +++ b/src/amd/common/ac_llvm_build.c
> @@ -554,21 +554,28 @@ LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
>   		chan[num_channels++] = LLVMGetUndef(elemtype);
>   
>   	return ac_build_gather_values(ctx, chan, 4);
>   }
>   
>   LLVMValueRef
>   ac_build_fdiv(struct ac_llvm_context *ctx,
>   	      LLVMValueRef num,
>   	      LLVMValueRef den)
>   {
> -	LLVMValueRef ret = LLVMBuildFDiv(ctx->builder, num, den, "");
> +	/* If we do (num / den), LLVM >= 7.0 does:
> +	 *    return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f : 1.0f));
> +	 *
> +	 * If we do (num * (1 / den)), LLVM does:
> +	 *    return num * v_rcp_f32(den);
> +	 */
> +	LLVMValueRef rcp = LLVMBuildFDiv(ctx->builder, ctx->f32_1, den, "");
> +	LLVMValueRef ret = LLVMBuildFMul(ctx->builder, num, rcp, "");
>   
>   	/* Use v_rcp_f32 instead of precise division. */
>   	if (!LLVMIsConstant(ret))
>   		LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
>   	return ret;
>   }
>   
>   /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
>    * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
>    * already multiplied by two. id is the cube face number.
>