[Mesa-dev] [PATCH 4/5] amd/common: save an instruction in the build_cube_select sequence

[Nicolai Hähnle]

From: Nicolai Hähnle <nicolai.haehnle at amd.com>

Avoid a v_cndmask: the absolute value is free due to input modifiers.
---
 src/amd/common/ac_llvm_build.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 8c050f31a76..71468df2dbc 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -431,26 +431,27 @@ build_cube_intrinsic(struct ac_llvm_context *ctx,
 
 /**
  * Build a manual selection sequence for cube face sc/tc coordinates and
  * major axis vector (multiplied by 2 for consistency) for the given
  * vec3 \p coords, for the face implied by \p selcoords.
  *
  * For the major axis, we always adjust the sign to be in the direction of
  * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
  * the selcoords major axis.
  */
-static void build_cube_select(LLVMBuilderRef builder,
+static void build_cube_select(struct ac_llvm_context *ctx,
 			      const struct cube_selection_coords *selcoords,
 			      const LLVMValueRef *coords,
 			      LLVMValueRef *out_st,
 			      LLVMValueRef *out_ma)
 {
+	LLVMBuilderRef builder = ctx->builder;
 	LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
 	LLVMValueRef is_ma_positive;
 	LLVMValueRef sgn_ma;
 	LLVMValueRef is_ma_z, is_not_ma_z;
 	LLVMValueRef is_ma_y;
 	LLVMValueRef is_ma_x;
 	LLVMValueRef sgn;
 	LLVMValueRef tmp;
 
 	is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
@@ -473,23 +474,23 @@ static void build_cube_select(LLVMBuilderRef builder,
 
 	/* Select tc */
 	tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
 	sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
 		LLVMConstReal(f32, -1.0), "");
 	out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
 
 	/* Select ma */
 	tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
 		LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
-	sgn = LLVMBuildSelect(builder, is_ma_positive,
-		LLVMConstReal(f32, 2.0), LLVMConstReal(f32, -2.0), "");
-	*out_ma = LLVMBuildFMul(builder, tmp, sgn, "");
+	tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
+				 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
+	*out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
 }
 
 void
 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 		       bool is_deriv, bool is_array, bool is_lod,
 		       LLVMValueRef *coords_arg,
 		       LLVMValueRef *derivs_arg)
 {
 
 	LLVMBuilderRef builder = ctx->builder;
@@ -563,21 +564,21 @@ ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 			 *            = 1/z * dx/dh - x/z * 1/z * dz/dh.
 			 *
 			 * This motivatives the implementation below.
 			 *
 			 * Whether this actually gives the expected results for
 			 * apps that might feed in derivatives obtained via
 			 * finite differences is anyone's guess. The OpenGL spec
 			 * seems awfully quiet about how textureGrad for cube
 			 * maps should be handled.
 			 */
-			build_cube_select(builder, &selcoords, &derivs_arg[axis * 3],
+			build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
 					  deriv_st, &deriv_ma);
 
 			deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
 
 			for (int i = 0; i < 2; ++i)
 				derivs[axis * 2 + i] =
 					LLVMBuildFSub(builder,
 						LLVMBuildFMul(builder, deriv_st[i], invma, ""),
 						LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
 		}
-- 
2.11.0