Merge pull request #433 from KhronosGroup/fix-428

Support returning arrays from functions in GLSL/MSL, refactor array copies in MSL.

reference/opt/shaders-hlsl/vert/return-array.vert

@@ -0,0 +1,31 @@
+static const float4 _20[2] = { 10.0f.xxxx, 20.0f.xxxx };
+
+static float4 gl_Position;
+static float4 vInput0;
+static float4 vInput1;
+
+struct SPIRV_Cross_Input
+{
+    float4 vInput0 : TEXCOORD0;
+    float4 vInput1 : TEXCOORD1;
+};
+
+struct SPIRV_Cross_Output
+{
+    float4 gl_Position : SV_Position;
+};
+
+void vert_main()
+{
+    gl_Position = 10.0f.xxxx + vInput1;
+}
+
+SPIRV_Cross_Output main(SPIRV_Cross_Input stage_input)
+{
+    vInput0 = stage_input.vInput0;
+    vInput1 = stage_input.vInput1;
+    vert_main();
+    SPIRV_Cross_Output stage_output;
+    stage_output.gl_Position = gl_Position;
+    return stage_output;
+}

reference/opt/shaders-msl/frag/constant-array.frag

@@ -1,3 +1,5 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+
 #include <metal_stdlib>
 #include <simd/simd.h>
 
@@ -9,6 +11,12 @@ struct Foobar
     float b;
 };
 
+constant float4 _37[3] = {float4(1.0), float4(2.0), float4(3.0)};
+constant float4 _49[2] = {float4(1.0), float4(2.0)};
+constant float4 _54[2] = {float4(8.0), float4(10.0)};
+constant float4 _55[2][2] = {{float4(1.0), float4(2.0)}, {float4(8.0), float4(10.0)}};
+constant Foobar _75[2] = {{10.0, 40.0}, {90.0, 70.0}};
+
 struct main0_in
 {
     int index [[user(locn0)]];
@@ -19,6 +27,20 @@ struct main0_out
     float4 FragColor [[color(0)]];
 };
 
+// Implementation of an array copy function to cover GLSL's ability to copy an array via assignment.
+template<typename T, uint N>
+void spvArrayCopy(thread T (&dst)[N], thread const T (&src)[N])
+{
+    for (uint i = 0; i < N; dst[i] = src[i], i++);
+}
+
+// An overload for constant arrays.
+template<typename T, uint N>
+void spvArrayCopyConstant(thread T (&dst)[N], constant T (&src)[N])
+{
+    for (uint i = 0; i < N; dst[i] = src[i], i++);
+}
+
 fragment main0_out main0(main0_in in [[stage_in]])
 {
     main0_out out = {};

reference/opt/shaders-msl/frag/constant-composites.frag

@@ -1,3 +1,5 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+
 #include <metal_stdlib>
 #include <simd/simd.h>
 
@@ -9,6 +11,9 @@ struct Foo
     float b;
 };
 
+constant float _16[4] = {1.0, 4.0, 3.0, 2.0};
+constant Foo _28[2] = {{10.0, 20.0}, {30.0, 40.0}};
+
 struct main0_in
 {
     int line [[user(locn0)]];
@@ -19,6 +24,20 @@ struct main0_out
     float4 FragColor [[color(0)]];
 };
 
+// Implementation of an array copy function to cover GLSL's ability to copy an array via assignment.
+template<typename T, uint N>
+void spvArrayCopy(thread T (&dst)[N], thread const T (&src)[N])
+{
+    for (uint i = 0; i < N; dst[i] = src[i], i++);
+}
+
+// An overload for constant arrays.
+template<typename T, uint N>
+void spvArrayCopyConstant(thread T (&dst)[N], constant T (&src)[N])
+{
+    for (uint i = 0; i < N; dst[i] = src[i], i++);
+}
+
 fragment main0_out main0(main0_in in [[stage_in]])
 {
     main0_out out = {};

reference/opt/shaders-msl/vert/return-array.vert

@@ -0,0 +1,24 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+constant float4 _20[2] = {float4(10.0), float4(20.0)};
+
+struct main0_in
+{
+    float4 vInput1 [[attribute(1)]];
+};
+
+struct main0_out
+{
+    float4 gl_Position [[position]];
+};
+
+vertex main0_out main0(main0_in in [[stage_in]])
+{
+    main0_out out = {};
+    out.gl_Position = float4(10.0) + in.vInput1;
+    return out;
+}
+

reference/opt/shaders/vert/return-array.vert

@@ -0,0 +1,9 @@
+#version 310 es
+
+layout(location = 1) in vec4 vInput1;
+
+void main()
+{
+    gl_Position = vec4(10.0) + vInput1;
+}
+

reference/shaders-hlsl/vert/return-array.vert

@@ -0,0 +1,48 @@
+static const float4 _20[2] = { 10.0f.xxxx, 20.0f.xxxx };
+
+static float4 gl_Position;
+static float4 vInput0;
+static float4 vInput1;
+
+struct SPIRV_Cross_Input
+{
+    float4 vInput0 : TEXCOORD0;
+    float4 vInput1 : TEXCOORD1;
+};
+
+struct SPIRV_Cross_Output
+{
+    float4 gl_Position : SV_Position;
+};
+
+void test(out float4 SPIRV_Cross_return_value[2])
+{
+    SPIRV_Cross_return_value = _20;
+}
+
+void test2(out float4 SPIRV_Cross_return_value[2])
+{
+    float4 foobar[2];
+    foobar[0] = vInput0;
+    foobar[1] = vInput1;
+    SPIRV_Cross_return_value = foobar;
+}
+
+void vert_main()
+{
+    float4 _42[2];
+    test(_42);
+    float4 _44[2];
+    test2(_44);
+    gl_Position = _42[0] + _44[1];
+}
+
+SPIRV_Cross_Output main(SPIRV_Cross_Input stage_input)
+{
+    vInput0 = stage_input.vInput0;
+    vInput1 = stage_input.vInput1;
+    vert_main();
+    SPIRV_Cross_Output stage_output;
+    stage_output.gl_Position = gl_Position;
+    return stage_output;
+}

reference/shaders-msl/asm/frag/op-constant-null.asm.frag

@@ -1,3 +1,5 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+
 #include <metal_stdlib>
 #include <simd/simd.h>
 
@@ -9,11 +11,27 @@ struct D
     float b;
 };
 
+constant float4 _14[4] = {float4(0.0), float4(0.0), float4(0.0), float4(0.0)};
+
 struct main0_out
 {
     float FragColor [[color(0)]];
 };
 
+// Implementation of an array copy function to cover GLSL's ability to copy an array via assignment.
+template<typename T, uint N>
+void spvArrayCopy(thread T (&dst)[N], thread const T (&src)[N])
+{
+    for (uint i = 0; i < N; dst[i] = src[i], i++);
+}
+
+// An overload for constant arrays.
+template<typename T, uint N>
+void spvArrayCopyConstant(thread T (&dst)[N], constant T (&src)[N])
+{
+    for (uint i = 0; i < N; dst[i] = src[i], i++);
+}
+
 fragment main0_out main0()
 {
     main0_out out = {};

reference/shaders-msl/frag/constant-array.frag

@@ -11,6 +11,12 @@ struct Foobar
     float b;
 };
 
+constant float4 _37[3] = {float4(1.0), float4(2.0), float4(3.0)};
+constant float4 _49[2] = {float4(1.0), float4(2.0)};
+constant float4 _54[2] = {float4(8.0), float4(10.0)};
+constant float4 _55[2][2] = {{float4(1.0), float4(2.0)}, {float4(8.0), float4(10.0)}};
+constant Foobar _75[2] = {{10.0, 40.0}, {90.0, 70.0}};
+
 struct main0_in
 {
     int index [[user(locn0)]];
@@ -21,6 +27,20 @@ struct main0_out
     float4 FragColor [[color(0)]];
 };
 
+// Implementation of an array copy function to cover GLSL's ability to copy an array via assignment.
+template<typename T, uint N>
+void spvArrayCopy(thread T (&dst)[N], thread const T (&src)[N])
+{
+    for (uint i = 0; i < N; dst[i] = src[i], i++);
+}
+
+// An overload for constant arrays.
+template<typename T, uint N>
+void spvArrayCopyConstant(thread T (&dst)[N], constant T (&src)[N])
+{
+    for (uint i = 0; i < N; dst[i] = src[i], i++);
+}
+
 float4 resolve(thread const Foobar& f)
 {
     return float4(f.a + f.b);

reference/shaders-msl/frag/constant-composites.frag

@@ -1,3 +1,5 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+
 #include <metal_stdlib>
 #include <simd/simd.h>
 
@@ -9,6 +11,9 @@ struct Foo
     float b;
 };
 
+constant float _16[4] = {1.0, 4.0, 3.0, 2.0};
+constant Foo _28[2] = {{10.0, 20.0}, {30.0, 40.0}};
+
 struct main0_in
 {
     int line [[user(locn0)]];
@@ -19,6 +24,20 @@ struct main0_out
     float4 FragColor [[color(0)]];
 };
 
+// Implementation of an array copy function to cover GLSL's ability to copy an array via assignment.
+template<typename T, uint N>
+void spvArrayCopy(thread T (&dst)[N], thread const T (&src)[N])
+{
+    for (uint i = 0; i < N; dst[i] = src[i], i++);
+}
+
+// An overload for constant arrays.
+template<typename T, uint N>
+void spvArrayCopyConstant(thread T (&dst)[N], constant T (&src)[N])
+{
+    for (uint i = 0; i < N; dst[i] = src[i], i++);
+}
+
 fragment main0_out main0(main0_in in [[stage_in]])
 {
     main0_out out = {};

reference/shaders-msl/vert/return-array.vert

@@ -0,0 +1,58 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+constant float4 _20[2] = {float4(10.0), float4(20.0)};
+
+struct main0_in
+{
+    float4 vInput1 [[attribute(1)]];
+    float4 vInput0 [[attribute(0)]];
+};
+
+struct main0_out
+{
+    float4 gl_Position [[position]];
+};
+
+// Implementation of an array copy function to cover GLSL's ability to copy an array via assignment.
+template<typename T, uint N>
+void spvArrayCopy(thread T (&dst)[N], thread const T (&src)[N])
+{
+    for (uint i = 0; i < N; dst[i] = src[i], i++);
+}
+
+// An overload for constant arrays.
+template<typename T, uint N>
+void spvArrayCopyConstant(thread T (&dst)[N], constant T (&src)[N])
+{
+    for (uint i = 0; i < N; dst[i] = src[i], i++);
+}
+
+void test(thread float4 (&SPIRV_Cross_return_value)[2])
+{
+    spvArrayCopyConstant(SPIRV_Cross_return_value, _20);
+}
+
+void test2(thread float4 (&SPIRV_Cross_return_value)[2], thread float4& vInput0, thread float4& vInput1)
+{
+    float4 foobar[2];
+    foobar[0] = vInput0;
+    foobar[1] = vInput1;
+    spvArrayCopy(SPIRV_Cross_return_value, foobar);
+}
+
+vertex main0_out main0(main0_in in [[stage_in]])
+{
+    main0_out out = {};
+    float4 _42[2];
+    test(_42);
+    float4 _44[2];
+    test2(_44, in.vInput0, in.vInput1);
+    out.gl_Position = _42[0] + _44[1];
+    return out;
+}
+

reference/shaders/vert/return-array.vert

@@ -0,0 +1,23 @@
+#version 310 es
+
+layout(location = 0) in vec4 vInput0;
+layout(location = 1) in vec4 vInput1;
+
+vec4[2] test()
+{
+    return vec4[](vec4(10.0), vec4(20.0));
+}
+
+vec4[2] test2()
+{
+    vec4 foobar[2];
+    foobar[0] = vInput0;
+    foobar[1] = vInput1;
+    return foobar;
+}
+
+void main()
+{
+    gl_Position = test()[0] + test2()[1];
+}
+

shaders-hlsl/vert/return-array.vert

@@ -0,0 +1,22 @@
+#version 310 es
+
+layout(location = 0) in vec4 vInput0;
+layout(location = 1) in vec4 vInput1;
+
+vec4[2] test()
+{
+	return vec4[](vec4(10.0), vec4(20.0));
+}
+
+vec4[2] test2()
+{
+	vec4 foobar[2];
+	foobar[0] = vInput0;
+	foobar[1] = vInput1;
+	return foobar;
+}
+
+void main()
+{
+	gl_Position = test()[0] + test2()[1];
+}

shaders-msl/vert/return-array.vert

@@ -0,0 +1,22 @@
+#version 310 es
+
+layout(location = 0) in vec4 vInput0;
+layout(location = 1) in vec4 vInput1;
+
+vec4[2] test()
+{
+	return vec4[](vec4(10.0), vec4(20.0));
+}
+
+vec4[2] test2()
+{
+	vec4 foobar[2];
+	foobar[0] = vInput0;
+	foobar[1] = vInput1;
+	return foobar;
+}
+
+void main()
+{
+	gl_Position = test()[0] + test2()[1];
+}

shaders/vert/return-array.vert

@@ -0,0 +1,22 @@
+#version 310 es
+
+layout(location = 0) in vec4 vInput0;
+layout(location = 1) in vec4 vInput1;
+
+vec4[2] test()
+{
+	return vec4[](vec4(10.0), vec4(20.0));
+}
+
+vec4[2] test2()
+{
+	vec4 foobar[2];
+	foobar[0] = vInput0;
+	foobar[1] = vInput1;
+	return foobar;
+}
+
+void main()
+{
+	gl_Position = test()[0] + test2()[1];
+}

spirv_glsl.cpp

@@ -5367,9 +5367,11 @@ void CompilerGLSL::emit_instruction(const Instruction &instruction)
 		length -= 3;
 
 		auto &callee = get<SPIRFunction>(func);
+		auto &return_type = get<SPIRType>(callee.return_type);
 		bool pure = function_is_pure(callee);
 
 		bool callee_has_out_variables = false;
+		bool emit_return_value_as_argument = false;
 
 		// Invalidate out variables passed to functions since they can be OpStore'd to.
 		for (uint32_t i = 0; i < length; i++)
@@ -5383,12 +5385,25 @@ void CompilerGLSL::emit_instruction(const Instruction &instruction)
 			flush_variable_declaration(arg[i]);
 		}
 
+		if (!return_type.array.empty() && !backend.can_return_array)
+		{
+			callee_has_out_variables = true;
+			emit_return_value_as_argument = true;
+		}
+
 		if (!pure)
 			register_impure_function_call();
 
 		string funexpr;
 		vector<string> arglist;
 		funexpr += to_name(func) + "(";
+
+		if (emit_return_value_as_argument)
+		{
+			statement(type_to_glsl(return_type), " ", to_name(id), type_to_array_glsl(return_type), ";");
+			arglist.push_back(to_name(id));
+		}
+
 		for (uint32_t i = 0; i < length; i++)
 		{
 			// Do not pass in separate images or samplers if we're remapping
@@ -5423,7 +5438,7 @@ void CompilerGLSL::emit_instruction(const Instruction &instruction)
 		// Check for function call constraints.
 		check_function_call_constraints(arg, length);
 
-		if (get<SPIRType>(result_type).basetype != SPIRType::Void)
+		if (return_type.basetype != SPIRType::Void)
 		{
 			// If the function actually writes to an out variable,
 			// take the conservative route and do not forward.
@@ -5435,7 +5450,13 @@ void CompilerGLSL::emit_instruction(const Instruction &instruction)
 			bool forward = args_will_forward(id, arg, length, pure) && !callee_has_out_variables && pure &&
 			               (forced_temporaries.find(id) == end(forced_temporaries));
 
-			emit_op(result_type, id, funexpr, forward);
+			if (emit_return_value_as_argument)
+			{
+				statement(funexpr, ";");
+				set<SPIRExpression>(id, to_name(id), result_type, true);
+			}
+			else
+				emit_op(result_type, id, funexpr, forward);
 
 			// Function calls are implicit loads from all variables in question.
 			// Set dependencies for them.
@@ -7645,6 +7666,7 @@ void CompilerGLSL::emit_function_prototype(SPIRFunction &func, uint64_t return_f
 	auto &type = get<SPIRType>(func.return_type);
 	decl += flags_to_precision_qualifiers_glsl(type, return_flags);
 	decl += type_to_glsl(type);
+	decl += type_to_array_glsl(type);
 	decl += " ";
 
 	if (func.self == entry_point)
@@ -8453,9 +8475,26 @@ void CompilerGLSL::emit_block_chain(SPIRBlock &block)
 
 		if (block.return_value)
 		{
-			// OpReturnValue can return Undef, so don't emit anything for this case.
+			auto &type = expression_type(block.return_value);
-			if (ids.at(block.return_value).get_type() != TypeUndef)
+			if (!type.array.empty() && !backend.can_return_array)
-				statement("return ", to_expression(block.return_value), ";");
+			{
+				// If we cannot return arrays, we will have a special out argument we can write to instead.
+				// The backend is responsible for setting this up, and redirection the return values as appropriate.
+				if (ids.at(block.return_value).get_type() != TypeUndef)
+					emit_array_copy("SPIRV_Cross_return_value", block.return_value);
+
+				if (!block_is_outside_flow_control_from_block(get<SPIRBlock>(current_function->entry_block), block) ||
+				    block.loop_dominator != SPIRBlock::NoDominator)
+				{
+					statement("return;");
+				}
+			}
+			else
+			{
+				// OpReturnValue can return Undef, so don't emit anything for this case.
+				if (ids.at(block.return_value).get_type() != TypeUndef)
+					statement("return ", to_expression(block.return_value), ";");
+			}
 		}
 		// If this block is the very final block and not called from control flow,
 		// we do not need an explicit return which looks out of place. Just end the function here.
@@ -8463,7 +8502,9 @@ void CompilerGLSL::emit_block_chain(SPIRBlock &block)
 		// but we actually need a return here ...
 		else if (!block_is_outside_flow_control_from_block(get<SPIRBlock>(current_function->entry_block), block) ||
 		         block.loop_dominator != SPIRBlock::NoDominator)
+		{
 			statement("return;");
+		}
 		break;
 
 	case SPIRBlock::Kill:
@@ -8581,3 +8622,8 @@ uint32_t CompilerGLSL::mask_relevant_memory_semantics(uint32_t semantics)
 	                    MemorySemanticsWorkgroupMemoryMask | MemorySemanticsUniformMemoryMask |
 	                    MemorySemanticsCrossWorkgroupMemoryMask | MemorySemanticsSubgroupMemoryMask);
 }
+
+void CompilerGLSL::emit_array_copy(const string &lhs, uint32_t rhs_id)
+{
+	statement(lhs, " = ", to_expression(rhs_id), ";");
+}

spirv_glsl.hpp

@@ -331,6 +331,7 @@ protected:
 		bool allow_precision_qualifiers = false;
 		bool can_swizzle_scalar = false;
 		bool force_gl_in_out_block = false;
+		bool can_return_array = true;
 	} backend;
 
 	void emit_struct(SPIRType &type);
@@ -424,6 +425,7 @@ protected:
 	std::string layout_for_variable(const SPIRVariable &variable);
 	std::string to_combined_image_sampler(uint32_t image_id, uint32_t samp_id);
 	virtual bool skip_argument(uint32_t id) const;
+	virtual void emit_array_copy(const std::string &lhs, uint32_t rhs_id);
 
 	bool buffer_is_packing_standard(const SPIRType &type, BufferPackingStandard packing, uint32_t start_offset = 0,
 	                                uint32_t end_offset = std::numeric_limits<uint32_t>::max());

spirv_hlsl.cpp

@@ -1663,9 +1663,17 @@ void CompilerHLSL::emit_function_prototype(SPIRFunction &func, uint64_t return_f
 	string decl;
 
 	auto &type = get<SPIRType>(func.return_type);
-	decl += flags_to_precision_qualifiers_glsl(type, return_flags);
+	if (type.array.empty())
-	decl += type_to_glsl(type);
+	{
-	decl += " ";
+		decl += flags_to_precision_qualifiers_glsl(type, return_flags);
+		decl += type_to_glsl(type);
+		decl += " ";
+	}
+	else
+	{
+		// We cannot return arrays in HLSL, so "return" through an out variable.
+		decl = "void ";
+	}
 
 	if (func.self == entry_point)
 	{
@@ -1683,6 +1691,19 @@ void CompilerHLSL::emit_function_prototype(SPIRFunction &func, uint64_t return_f
 		decl += to_name(func.self);
 
 	decl += "(";
+
+	if (!type.array.empty())
+	{
+		// Fake array returns by writing to an out array instead.
+		decl += "out ";
+		decl += type_to_glsl(type);
+		decl += " ";
+		decl += "SPIRV_Cross_return_value";
+		decl += type_to_array_glsl(type);
+		if (!func.arguments.empty())
+			decl += ", ";
+	}
+
 	for (auto &arg : func.arguments)
 	{
 		// Might change the variable name if it already exists in this function.
@@ -3802,6 +3823,7 @@ string CompilerHLSL::compile()
 	backend.can_swizzle_scalar = true;
 	backend.can_declare_struct_inline = false;
 	backend.can_declare_arrays_inline = false;
+	backend.can_return_array = false;
 
 	update_active_builtins();
 	analyze_sampler_comparison_states();

spirv_msl.cpp

@@ -72,6 +72,8 @@ string CompilerMSL::compile()
 	backend.use_typed_initializer_list = true;
 	backend.native_row_major_matrix = false;
 	backend.flexible_member_array_supported = false;
+	backend.can_declare_arrays_inline = false;
+	backend.can_return_array = false;
 
 	replace_illegal_names();
 
@@ -912,11 +914,19 @@ void CompilerMSL::emit_custom_functions()
 
 		case SPVFuncImplArrayCopy:
 			statement("// Implementation of an array copy function to cover GLSL's ability to copy an array via "
-			          "assignment. ");
+			          "assignment.");
-			statement("template<typename T>");
+			statement("template<typename T, uint N>");
-			statement("void spvArrayCopy(thread T* dst, thread const T* src, uint count)");
+			statement("void spvArrayCopy(thread T (&dst)[N], thread const T (&src)[N])");
 			begin_scope();
-			statement("for (uint i = 0; i < count; *dst++ = *src++, i++);");
+			statement("for (uint i = 0; i < N; dst[i] = src[i], i++);");
+			end_scope();
+			statement("");
+
+			statement("// An overload for constant arrays.");
+			statement("template<typename T, uint N>");
+			statement("void spvArrayCopyConstant(thread T (&dst)[N], constant T (&src)[N])");
+			begin_scope();
+			statement("for (uint i = 0; i < N; dst[i] = src[i], i++);");
 			end_scope();
 			statement("");
 			break;
@@ -1134,6 +1144,34 @@ void CompilerMSL::declare_undefined_values()
 		statement("");
 }
 
+void CompilerMSL::declare_constant_arrays()
+{
+	// MSL cannot declare arrays inline (except when declaring a variable), so we must move them out to
+	// global constants directly, so we are able to use constants as variable expressions.
+	bool emitted = false;
+
+	for (auto &id : ids)
+	{
+		if (id.get_type() == TypeConstant)
+		{
+			auto &c = id.get<SPIRConstant>();
+			if (c.specialization)
+				continue;
+
+			auto &type = get<SPIRType>(c.constant_type);
+			if (!type.array.empty())
+			{
+				auto name = to_name(c.self);
+				statement("constant ", variable_decl(type, name), " = ", constant_expression(c), ";");
+				emitted = true;
+			}
+		}
+	}
+
+	if (emitted)
+		statement("");
+}
+
 void CompilerMSL::emit_resources()
 {
 	// Output non-interface structs. These include local function structs
@@ -1170,6 +1208,7 @@ void CompilerMSL::emit_resources()
 		}
 	}
 
+	declare_constant_arrays();
 	declare_undefined_values();
 
 	// Output interface structs.
@@ -1743,27 +1782,44 @@ bool CompilerMSL::maybe_emit_input_struct_assignment(uint32_t id_lhs, uint32_t i
 	return true;
 }
 
+void CompilerMSL::emit_array_copy(const string &lhs, uint32_t rhs_id)
+{
+	// Assignment from an array initializer is fine.
+	if (ids[rhs_id].get_type() == TypeConstant)
+		statement("spvArrayCopyConstant(", lhs, ", ", to_expression(rhs_id), ");");
+	else
+		statement("spvArrayCopy(", lhs, ", ", to_expression(rhs_id), ");");
+}
+
 // Since MSL does not allow arrays to be copied via simple variable assignment,
 // if the LHS and RHS represent an assignment of an entire array, it must be
 // implemented by calling an array copy function.
 // Returns whether the struct assignment was emitted.
 bool CompilerMSL::maybe_emit_array_assignment(uint32_t id_lhs, uint32_t id_rhs)
 {
-	// Assignment from an array initializer is fine.
-	if (ids[id_rhs].get_type() == TypeConstant)
-		return false;
-
 	// We only care about assignments of an entire array
 	auto &type = expression_type(id_rhs);
 	if (type.array.size() == 0)
 		return false;
 
+	auto *var = maybe_get<SPIRVariable>(id_lhs);
+	if (ids[id_rhs].get_type() == TypeConstant && var && var->deferred_declaration)
+	{
+		// Special case, if we end up declaring a variable when assigning the constant array,
+		// we can avoid the copy by directly assigning the constant expression.
+		// This is likely necessary to be able to use a variable as a true look-up table, as it is unlikely
+		// the compiler will be able to optimize the spvArrayCopy() into a constant LUT.
+		// After a variable has been declared, we can no longer assign constant arrays in MSL unfortunately.
+		statement(to_expression(id_lhs), " = ", constant_expression(get<SPIRConstant>(id_rhs)), ";");
+		return true;
+	}
+
 	// Ensure the LHS variable has been declared
 	auto *p_v_lhs = maybe_get_backing_variable(id_lhs);
 	if (p_v_lhs)
 		flush_variable_declaration(p_v_lhs->self);
 
-	statement("spvArrayCopy(", to_expression(id_lhs), ", ", to_expression(id_rhs), ", ", to_array_size(type, 0), ");");
+	emit_array_copy(to_expression(id_lhs), id_rhs);
 	register_write(id_lhs);
 
 	return true;
@@ -1949,12 +2005,32 @@ void CompilerMSL::emit_function_prototype(SPIRFunction &func, uint64_t)
 	processing_entry_point = (func.self == entry_point);
 
 	auto &type = get<SPIRType>(func.return_type);
-	decl += func_type_decl(type);
+
+	if (type.array.empty())
+	{
+		decl += func_type_decl(type);
+	}
+	else
+	{
+		// We cannot return arrays in MSL, so "return" through an out variable.
+		decl = "void";
+	}
+
 	decl += " ";
 	decl += to_name(func.self);
-
 	decl += "(";
 
+	if (!type.array.empty())
+	{
+		// Fake arrays returns by writing to an out array instead.
+		decl += "thread ";
+		decl += type_to_glsl(type);
+		decl += " (&SPIRV_Cross_return_value)";
+		decl += type_to_array_glsl(type);
+		if (!func.arguments.empty())
+			decl += ", ";
+	}
+
 	if (processing_entry_point)
 	{
 		decl += entry_point_args(!func.arguments.empty());
@@ -2595,7 +2671,7 @@ string CompilerMSL::func_type_decl(SPIRType &type)
 {
 	auto &execution = get_entry_point();
 	// The regular function return type. If not processing the entry point function, that's all we need
-	string return_type = type_to_glsl(type);
+	string return_type = type_to_glsl(type) + type_to_array_glsl(type);
 	if (!processing_entry_point)
 		return return_type;
 
@@ -2604,7 +2680,7 @@ string CompilerMSL::func_type_decl(SPIRType &type)
 	{
 		auto &so_var = get<SPIRVariable>(stage_out_var_id);
 		auto &so_type = get<SPIRType>(so_var.basetype);
-		return_type = type_to_glsl(so_type);
+		return_type = type_to_glsl(so_type) + type_to_array_glsl(type);
 	}
 
 	// Prepend a entry type, based on the execution model
@@ -3522,16 +3598,38 @@ CompilerMSL::SPVFuncImpl CompilerMSL::OpCodePreprocessor::get_spv_func_impl(Op o
 	case OpFMod:
 		return SPVFuncImplMod;
 
+	case OpFunctionCall:
+	{
+		auto &return_type = compiler.get<SPIRType>(args[0]);
+		if (!return_type.array.empty())
+			return SPVFuncImplArrayCopy;
+		else
+			return SPVFuncImplNone;
+	}
+
 	case OpStore:
 	{
 		// Get the result type of the RHS. Since this is run as a pre-processing stage,
 		// we must extract the result type directly from the Instruction, rather than the ID.
 		uint32_t id_rhs = args[1];
-		uint32_t type_id_rhs = result_types[id_rhs];
-		if ((compiler.ids[id_rhs].get_type() != TypeConstant) && type_id_rhs &&
-		    compiler.is_array(compiler.get<SPIRType>(type_id_rhs)))
-			return SPVFuncImplArrayCopy;
 
+		const SPIRType *type = nullptr;
+		if (compiler.ids[id_rhs].get_type() != TypeNone)
+		{
+			// Could be a constant, or similar.
+			type = &compiler.expression_type(id_rhs);
+		}
+		else
+		{
+			// Or ... an expression.
+			if (result_types[id_rhs] != 0)
+				type = &compiler.get<SPIRType>(result_types[id_rhs]);
+		}
+
+		if (type && compiler.is_array(*type))
+			return SPVFuncImplArrayCopy;
+		else
+			return SPVFuncImplNone;
 		break;
 	}
 

spirv_msl.hpp

@@ -208,6 +208,7 @@ protected:
 	std::string to_qualifiers_glsl(uint32_t id) override;
 	void replace_illegal_names() override;
 	void declare_undefined_values() override;
+	void declare_constant_arrays();
 	bool is_non_native_row_major_matrix(uint32_t id) override;
 	bool member_is_non_native_row_major_matrix(const SPIRType &type, uint32_t index) override;
 	std::string convert_row_major_matrix(std::string exp_str, const SPIRType &exp_type) override;
@@ -265,6 +266,7 @@ protected:
 	const char *get_memory_order(uint32_t spv_mem_sem);
 	void add_pragma_line(const std::string &line);
 	void emit_barrier(uint32_t id_exe_scope, uint32_t id_mem_scope, uint32_t id_mem_sem);
+	void emit_array_copy(const std::string &lhs, uint32_t rhs_id) override;
 
 	Options options;
 	std::set<SPVFuncImpl> spv_function_implementations;