Merge pull request #2430 from KhronosGroup/hlsl-mesh-single-clip-cull…

…-fix

HLSL: Fix clip/cull lowering for single element.

reference/shaders-hlsl-no-opt/mesh/clip-cull-array-load-store-single.sm65.nofxc.fxconly.spv16.mesh

@@ -0,0 +1,36 @@
+static uint gl_LocalInvocationIndex;
+struct SPIRV_Cross_Input
+{
+    uint gl_LocalInvocationIndex : SV_GroupIndex;
+};
+
+struct gl_MeshPerVertexEXT
+{
+    float gl_ClipDistance[1] : SV_ClipDistance;
+};
+
+struct gl_MeshPerPrimitiveEXT
+{
+};
+
+void write_clip_distance(inout float v[1])
+{
+    v[0] += 1.0f;
+}
+
+void mesh_main(out gl_MeshPerVertexEXT gl_MeshVerticesEXT[3])
+{
+    SetMeshOutputCounts(3u, 1u);
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance[0] = 4.0f;
+    float param[1] = gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance;
+    write_clip_distance(param);
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance = param;
+}
+
+[outputtopology("triangle")]
+[numthreads(1, 1, 1)]
+void main(SPIRV_Cross_Input stage_input, out vertices gl_MeshPerVertexEXT gl_MeshVerticesEXT[3])
+{
+    gl_LocalInvocationIndex = stage_input.gl_LocalInvocationIndex;
+    mesh_main(gl_MeshVerticesEXT);
+}

reference/shaders-hlsl-no-opt/mesh/clip-cull-array-load-store.sm65.nofxc.fxconly.spv16.mesh

@@ -28,7 +28,7 @@ void mesh_main(out gl_MeshPerVertexEXT gl_MeshVerticesEXT[3])
     gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance[1] = 4.0f;
     gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance[2] = 4.0f;
     gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance[3] = 4.0f;
-    float _62[4] = { gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance[0], gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance[1], gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance[2], gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance[3] };
+    float _62[4] = { gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance.x, gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance.y, gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance.z, gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance.w };
     float param[4] = _62;
     write_clip_distance(param);
     gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance = float4(param[0], param[1], param[2], param[3]);

shaders-hlsl-no-opt/mesh/clip-cull-array-load-store-single.sm65.nofxc.fxconly.spv16.mesh

@@ -0,0 +1,20 @@
+#version 450
+#extension GL_EXT_mesh_shader : require
+layout(triangles, max_vertices = 3, max_primitives = 1) out;
+
+out gl_MeshPerVertexEXT
+{
+	float gl_ClipDistance[1];
+} gl_MeshVerticesEXT[];
+
+void write_clip_distance(inout float v[1])
+{
+	v[0] += 1.0;
+}
+
+void main()
+{
+	SetMeshOutputsEXT(3, 1);
+	gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance[0] = 4.0;
+	write_clip_distance(gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance);
+}

spirv_hlsl.cpp

@@ -4827,14 +4827,15 @@ void CompilerHLSL::emit_load(const Instruction &instruction)
 		    has_decoration(ptr, DecorationBuiltIn) &&
 		    (get_decoration(ptr, DecorationBuiltIn) == BuiltInClipDistance ||
 		     get_decoration(ptr, DecorationBuiltIn) == BuiltInCullDistance) &&
-		    is_array(res_type) && !is_array(get<SPIRType>(res_type.parent_type)))
+		    is_array(res_type) && !is_array(get<SPIRType>(res_type.parent_type)) &&
+		    to_array_size_literal(res_type) > 1)
 		{
 			track_expression_read(ptr);
 			string load_expr = "{ ";
 			uint32_t num_elements = to_array_size_literal(res_type);
 			for (uint32_t i = 0; i < num_elements; i++)
 			{
-				load_expr += join(to_expression(ptr), "[", i, "]");
+				load_expr += join(to_expression(ptr), ".", index_to_swizzle(i));
 				if (i + 1 < num_elements)
 					load_expr += ", ";
 			}
@@ -6960,6 +6961,13 @@ void CompilerHLSL::cast_to_variable_store(uint32_t target_id, std::string &expr,
 	if (num_clip > 4)
 		SPIRV_CROSS_THROW("Number of clip or cull distances exceeds 4, this will not work with mesh shaders.");
 
+	if (num_clip == 1)
+	{
+		// We already emit array here.
+		CompilerGLSL::cast_to_variable_store(target_id, expr, expr_type);
+		return;
+	}
+
 	auto unrolled_expr = join("float", num_clip, "(");
 	for (uint32_t i = 0; i < num_clip; i++)
 	{