Merge pull request #648 from billhollings/master

CompilerMSL disable rasterization on buffer writes in vertex shader.

reference/opt/shaders-msl/vert/no_stage_out.write_buff.vert

@@ -0,0 +1,37 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+struct _35
+{
+    uint4 _m0[1024];
+};
+
+struct _40
+{
+    uint4 _m0[1024];
+};
+
+struct main0_out
+{
+    float4 gl_Position [[position]];
+};
+
+struct main0_in
+{
+    float4 m_17 [[attribute(0)]];
+};
+
+vertex void main0(main0_in in [[stage_in]], constant _40& _42 [[buffer(0)]], device _35& _37 [[buffer(1)]])
+{
+    main0_out out = {};
+    out.gl_Position = in.m_17;
+    for (int _51 = 0; _51 < 1024; )
+    {
+        _37._m0[_51] = _42._m0[_51];
+        _51++;
+        continue;
+    }
+}
+

reference/opt/shaders-msl/vert/no_stage_out.write_buff_atomic.vert

@@ -0,0 +1,30 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+
+#include <metal_stdlib>
+#include <simd/simd.h>
+#include <metal_atomic>
+
+using namespace metal;
+
+struct _23
+{
+    uint _m0;
+};
+
+struct main0_out
+{
+    float4 gl_Position [[position]];
+};
+
+struct main0_in
+{
+    float4 m_17 [[attribute(0)]];
+};
+
+vertex void main0(main0_in in [[stage_in]], device _23& _25 [[buffer(0)]])
+{
+    main0_out out = {};
+    out.gl_Position = in.m_17;
+    uint _29 = atomic_fetch_add_explicit((volatile device atomic_uint*)&_25._m0, 1u, memory_order_relaxed);
+}
+

reference/shaders-msl/vert/no_stage_out.write_buff.vert

@@ -0,0 +1,35 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+struct _35
+{
+    uint4 _m0[1024];
+};
+
+struct _40
+{
+    uint4 _m0[1024];
+};
+
+struct main0_out
+{
+    float4 gl_Position [[position]];
+};
+
+struct main0_in
+{
+    float4 m_17 [[attribute(0)]];
+};
+
+vertex void main0(main0_in in [[stage_in]], constant _40& _42 [[buffer(0)]], device _35& _37 [[buffer(1)]])
+{
+    main0_out out = {};
+    out.gl_Position = in.m_17;
+    for (int _22 = 0; _22 < 1024; _22++)
+    {
+        _37._m0[_22] = _42._m0[_22];
+    }
+}
+

reference/shaders-msl/vert/no_stage_out.write_buff_atomic.vert

@@ -0,0 +1,31 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+
+#include <metal_stdlib>
+#include <simd/simd.h>
+#include <metal_atomic>
+
+using namespace metal;
+
+struct _23
+{
+    uint _m0;
+};
+
+struct main0_out
+{
+    float4 gl_Position [[position]];
+};
+
+struct main0_in
+{
+    float4 m_17 [[attribute(0)]];
+};
+
+vertex void main0(main0_in in [[stage_in]], device _23& _25 [[buffer(0)]])
+{
+    main0_out out = {};
+    out.gl_Position = in.m_17;
+    uint _29 = atomic_fetch_add_explicit((volatile device atomic_uint*)&_25._m0, 1u, memory_order_relaxed);
+    uint _22 = _29;
+}
+

shaders-msl/vert/no_stage_out.write_buff.vert

@@ -0,0 +1,23 @@
+#version 450
+
+layout(binding = 1, std430) writeonly buffer _33_35
+{
+	uvec4 _m0[1024];
+} _35;
+
+layout(binding = 0, std140) uniform _38_40
+{
+	uvec4 _m0[1024];
+} _40;
+
+layout(location = 0) in vec4 _14;
+
+void main()
+{
+	gl_Position = _14;
+	for (int _19 = 0; _19 < 1024; _19++)
+	{
+		_35._m0[_19] = _40._m0[_19];
+	}
+}
+

shaders-msl/vert/no_stage_out.write_buff_atomic.vert

@@ -0,0 +1,15 @@
+#version 450
+
+layout(binding = 0, std430) coherent buffer _19_21
+{
+	uint _m0;
+} _21;
+
+layout(location = 0) in vec4 _14;
+
+void main()
+{
+	gl_Position = _14;
+	uint _26 = atomicAdd(_21._m0, 1u);
+}
+

spirv_msl.cpp

@@ -383,8 +383,8 @@ void CompilerMSL::preprocess_op_codes()
 		add_pragma_line("#pragma clang diagnostic ignored \"-Wunused-variable\"");
 	}
 
-	// Metal vertex functions that write to textures must disable rasterization and return void.
-	if (preproc.uses_image_write)
+	// Metal vertex functions that write to resources must disable rasterization and return void.
+	if (preproc.uses_resource_write)
 		is_rasterization_disabled = true;
 }
 
@@ -3141,8 +3141,8 @@ string CompilerMSL::get_argument_address_space(const SPIRVariable &argument)
 
 	case StorageClassStorageBuffer:
 	{
-		auto flags = get_buffer_block_flags(argument);
-		return flags.get(DecorationNonWritable) ? "const device" : "device";
+		bool readonly = get_buffer_block_flags(argument).get(DecorationNonWritable);
+		return readonly ? "const device" : "device";
 	}
 
 	case StorageClassUniform:
@@ -3151,13 +3151,13 @@ string CompilerMSL::get_argument_address_space(const SPIRVariable &argument)
 		if (type.basetype == SPIRType::Struct)
 		{
 			bool ssbo = has_decoration(type.self, DecorationBufferBlock);
-			if (!ssbo)
-				return "constant";
-			else
+			if (ssbo)
 			{
 				bool readonly = get_buffer_block_flags(argument).get(DecorationNonWritable);
 				return readonly ? "const device" : "device";
 			}
+			else
+				return "constant";
 		}
 		break;
 
@@ -4039,13 +4039,16 @@ bool CompilerMSL::OpCodePreprocessor::handle(Op opcode, const uint32_t *args, ui
 		break;
 
 	case OpImageWrite:
-		uses_image_write = true;
+		uses_resource_write = true;
+		break;
+
+	case OpStore:
+		check_resource_write(args[0]);
 		break;
 
 	case OpAtomicExchange:
 	case OpAtomicCompareExchange:
 	case OpAtomicCompareExchangeWeak:
-	case OpAtomicLoad:
 	case OpAtomicIIncrement:
 	case OpAtomicIDecrement:
 	case OpAtomicIAdd:
@@ -4057,6 +4060,11 @@ bool CompilerMSL::OpCodePreprocessor::handle(Op opcode, const uint32_t *args, ui
 	case OpAtomicAnd:
 	case OpAtomicOr:
 	case OpAtomicXor:
+		uses_atomics = true;
+		check_resource_write(args[2]);
+		break;
+
+	case OpAtomicLoad:
 		uses_atomics = true;
 		break;
 
@@ -4072,6 +4080,15 @@ bool CompilerMSL::OpCodePreprocessor::handle(Op opcode, const uint32_t *args, ui
 	return true;
 }
 
+// If the variable is a Uniform or StorageBuffer, mark that a resource has been written to.
+void CompilerMSL::OpCodePreprocessor::check_resource_write(uint32_t var_id)
+{
+	auto *p_var = compiler.maybe_get_backing_variable(var_id);
+	StorageClass sc = p_var ? p_var->storage : StorageClassMax;
+	if (sc == StorageClassUniform || sc == StorageClassStorageBuffer)
+		uses_resource_write = true;
+}
+
 // Returns an enumeration of a SPIR-V function that needs to be output for certain Op codes.
 CompilerMSL::SPVFuncImpl CompilerMSL::OpCodePreprocessor::get_spv_func_impl(Op opcode, const uint32_t *args)
 {

spirv_msl.hpp

@@ -403,12 +403,13 @@ protected:
 
 		bool handle(spv::Op opcode, const uint32_t *args, uint32_t length) override;
 		CompilerMSL::SPVFuncImpl get_spv_func_impl(spv::Op opcode, const uint32_t *args);
+		void check_resource_write(uint32_t var_id);
 
 		CompilerMSL &compiler;
 		std::unordered_map<uint32_t, uint32_t> result_types;
 		bool suppress_missing_prototypes = false;
 		bool uses_atomics = false;
-		bool uses_image_write = false;
+		bool uses_resource_write = false;
 	};
 
 	// Sorts the members of a SPIRType and associated Meta info based on a settable sorting