MSL: Reintroduce workaround for constant arrays being passed by value.

reference/shaders-msl-no-opt/vert/pass-array-by-value.force-native-array.vert

@@ -0,0 +1,103 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+constant float4 _68[4] = { float4(0.0), float4(1.0), float4(2.0), float4(3.0) };
+
+struct main0_out
+{
+    float4 gl_Position [[position]];
+};
+
+struct main0_in
+{
+    int Index1 [[attribute(0)]];
+    int Index2 [[attribute(1)]];
+};
+
+template<typename T, uint A>
+inline void spvArrayCopyFromConstantToStack1(thread T (&dst)[A], constant T (&src)[A])
+{
+    for (uint i = 0; i < A; i++)
+    {
+        dst[i] = src[i];
+    }
+}
+
+template<typename T, uint A>
+inline void spvArrayCopyFromConstantToThreadGroup1(threadgroup T (&dst)[A], constant T (&src)[A])
+{
+    for (uint i = 0; i < A; i++)
+    {
+        dst[i] = src[i];
+    }
+}
+
+template<typename T, uint A>
+inline void spvArrayCopyFromStackToStack1(thread T (&dst)[A], thread const T (&src)[A])
+{
+    for (uint i = 0; i < A; i++)
+    {
+        dst[i] = src[i];
+    }
+}
+
+template<typename T, uint A>
+inline void spvArrayCopyFromStackToThreadGroup1(threadgroup T (&dst)[A], thread const T (&src)[A])
+{
+    for (uint i = 0; i < A; i++)
+    {
+        dst[i] = src[i];
+    }
+}
+
+template<typename T, uint A>
+inline void spvArrayCopyFromThreadGroupToStack1(thread T (&dst)[A], threadgroup const T (&src)[A])
+{
+    for (uint i = 0; i < A; i++)
+    {
+        dst[i] = src[i];
+    }
+}
+
+template<typename T, uint A>
+inline void spvArrayCopyFromThreadGroupToThreadGroup1(threadgroup T (&dst)[A], threadgroup const T (&src)[A])
+{
+    for (uint i = 0; i < A; i++)
+    {
+        dst[i] = src[i];
+    }
+}
+
+static inline __attribute__((always_inline))
+float4 consume_constant_arrays2(thread const float4 (&positions)[4], thread const float4 (&positions2)[4], thread int& Index1, thread int& Index2)
+{
+    float4 indexable[4];
+    spvArrayCopyFromStackToStack1(indexable, positions);
+    float4 indexable_1[4];
+    spvArrayCopyFromStackToStack1(indexable_1, positions2);
+    return indexable[Index1] + indexable_1[Index2];
+}
+
+static inline __attribute__((always_inline))
+float4 consume_constant_arrays(thread const float4 (&positions)[4], thread const float4 (&positions2)[4], thread int& Index1, thread int& Index2)
+{
+    return consume_constant_arrays2(positions, positions2, Index1, Index2);
+}
+
+vertex main0_out main0(main0_in in [[stage_in]])
+{
+    float4 _68_array_copy[4] = { float4(0.0), float4(1.0), float4(2.0), float4(3.0) };
+    main0_out out = {};
+    float4 LUT2[4];
+    LUT2[0] = float4(10.0);
+    LUT2[1] = float4(11.0);
+    LUT2[2] = float4(12.0);
+    LUT2[3] = float4(13.0);
+    out.gl_Position = consume_constant_arrays(_68_array_copy, LUT2, in.Index1, in.Index2);
+    return out;
+}
+

shaders-msl-no-opt/vert/pass-array-by-value.force-native-array.vert

@@ -0,0 +1,26 @@
+#version 310 es
+
+layout(location = 0) in int Index1;
+layout(location = 1) in int Index2;
+
+vec4 consume_constant_arrays2(const vec4 positions[4], const vec4 positions2[4])
+{
+	return positions[Index1] + positions2[Index2];
+}
+
+vec4 consume_constant_arrays(const vec4 positions[4], const vec4 positions2[4])
+{
+	return consume_constant_arrays2(positions, positions2);
+}
+
+const vec4 LUT1[] = vec4[](vec4(0.0), vec4(1.0), vec4(2.0), vec4(3.0));
+
+void main()
+{
+	vec4 LUT2[4];
+	LUT2[0] = vec4(10.0);
+	LUT2[1] = vec4(11.0);
+	LUT2[2] = vec4(12.0);
+	LUT2[3] = vec4(13.0);
+	gl_Position = consume_constant_arrays(LUT1, LUT2);
+}

spirv_common.hpp

@@ -939,6 +939,11 @@ struct SPIRFunction : IVariant
 	// Intentionally not a small vector, this one is rare, and std::function can be large.
 	Vector<std::function<void()>> fixup_hooks_in;
 
+	// On function entry, make sure to copy a constant array into thread addr space to work around
+	// the case where we are passing a constant array by value to a function on backends which do not
+	// consider arrays value types.
+	SmallVector<ID> constant_arrays_needed_on_stack;
+
 	bool active = false;
 	bool flush_undeclared = true;
 	bool do_combined_parameters = true;

spirv_glsl.cpp

@@ -11751,6 +11751,14 @@ void CompilerGLSL::emit_function(SPIRFunction &func, const Bitset &return_flags)
 	current_function = &func;
 	auto &entry_block = get<SPIRBlock>(func.entry_block);
 
+	sort(begin(func.constant_arrays_needed_on_stack), end(func.constant_arrays_needed_on_stack));
+	for (auto &array : func.constant_arrays_needed_on_stack)
+	{
+		auto &c = get<SPIRConstant>(array);
+		auto &type = get<SPIRType>(c.constant_type);
+		statement(variable_decl(type, join("_", array, "_array_copy")), " = ", constant_expression(c), ";");
+	}
+
 	for (auto &v : func.local_variables)
 	{
 		auto &var = get<SPIRVariable>(v);

spirv_msl.cpp

@@ -8205,7 +8205,29 @@ string CompilerMSL::to_func_call_arg(const SPIRFunction::Parameter &arg, uint32_
 	if (is_dynamic_img_sampler && !arg_is_dynamic_img_sampler)
 		arg_str = join("spvDynamicImageSampler<", type_to_glsl(get<SPIRType>(type.image.type)), ">(");
 
-	arg_str += CompilerGLSL::to_func_call_arg(arg, id);
+	auto *c = maybe_get<SPIRConstant>(id);
+	if (msl_options.force_native_arrays && c && !get<SPIRType>(c->constant_type).array.empty())
+	{
+		// If we are passing a constant array directly to a function for some reason,
+		// the callee will expect an argument in thread const address space
+		// (since we can only bind to arrays with references in MSL).
+		// To resolve this, we must emit a copy in this address space.
+		// This kind of code gen should be rare enough that performance is not a real concern.
+		// Inline the SPIR-V to avoid this kind of suboptimal codegen.
+		//
+		// We risk calling this inside a continue block (invalid code),
+		// so just create a thread local copy in the current function.
+		arg_str = join("_", id, "_array_copy");
+		auto &constants = current_function->constant_arrays_needed_on_stack;
+		auto itr = find(begin(constants), end(constants), ID(id));
+		if (itr == end(constants))
+		{
+			force_recompile();
+			constants.push_back(id);
+		}
+	}
+	else
+		arg_str += CompilerGLSL::to_func_call_arg(arg, id);
 
 	// Need to check the base variable in case we need to apply a qualified alias.
 	uint32_t var_id = 0;
@@ -10029,10 +10051,34 @@ string CompilerMSL::argument_decl(const SPIRFunction::Parameter &arg)
 	    (storage == StorageClassFunction || storage == StorageClassGeneric))
 	{
 		// If the argument is a pure value and not an opaque type, we will pass by value.
-		if (!address_space.empty())
-			decl = join(address_space, " ", decl);
-		decl += " ";
-		decl += to_expression(name_id);
+		if (msl_options.force_native_arrays && is_array(type))
+		{
+			// We are receiving an array by value. This is problematic.
+			// We cannot be sure of the target address space since we are supposed to receive a copy,
+			// but this is not possible with MSL without some extra work.
+			// We will have to assume we're getting a reference in thread address space.
+			// If we happen to get a reference in constant address space, the caller must emit a copy and pass that.
+			// Thread const therefore becomes the only logical choice, since we cannot "create" a constant array from
+			// non-constant arrays, but we can create thread const from constant.
+			decl = string("thread const ") + decl;
+			decl += " (&";
+			const char *restrict_kw = to_restrict(name_id);
+			if (*restrict_kw)
+			{
+				decl += " ";
+				decl += restrict_kw;
+			}
+			decl += to_expression(name_id);
+			decl += ")";
+			decl += type_to_array_glsl(type);
+		}
+		else
+		{
+			if (!address_space.empty())
+				decl = join(address_space, " ", decl);
+			decl += " ";
+			decl += to_expression(name_id);
+		}
 	}
 	else if (is_array(type) && !type_is_image)
 	{