Updated spirv-cross.

3rdparty/spirv-cross/spirv.h

@@ -54,11 +54,11 @@
 typedef unsigned int SpvId;
 
 #define SPV_VERSION 0x10500
-#define SPV_REVISION 1
+#define SPV_REVISION 3
 
 static const unsigned int SpvMagicNumber = 0x07230203;
 static const unsigned int SpvVersion = 0x00010500;
-static const unsigned int SpvRevision = 1;
+static const unsigned int SpvRevision = 3;
 static const unsigned int SpvOpCodeMask = 0xffff;
 static const unsigned int SpvWordCountShift = 16;
 
@@ -1899,6 +1899,13 @@ inline void SpvHasResultAndType(SpvOp opcode, bool *hasResult, bool *hasResultTy
     case SpvOpSubgroupAnyKHR: *hasResult = true; *hasResultType = true; break;
     case SpvOpSubgroupAllEqualKHR: *hasResult = true; *hasResultType = true; break;
     case SpvOpSubgroupReadInvocationKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpTypeRayQueryProvisionalKHR: *hasResult = true; *hasResultType = false; break;
+    case SpvOpRayQueryInitializeKHR: *hasResult = false; *hasResultType = false; break;
+    case SpvOpRayQueryTerminateKHR: *hasResult = false; *hasResultType = false; break;
+    case SpvOpRayQueryGenerateIntersectionKHR: *hasResult = false; *hasResultType = false; break;
+    case SpvOpRayQueryConfirmIntersectionKHR: *hasResult = false; *hasResultType = false; break;
+    case SpvOpRayQueryProceedKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionTypeKHR: *hasResult = true; *hasResultType = true; break;
     case SpvOpGroupIAddNonUniformAMD: *hasResult = true; *hasResultType = true; break;
     case SpvOpGroupFAddNonUniformAMD: *hasResult = true; *hasResultType = true; break;
     case SpvOpGroupFMinNonUniformAMD: *hasResult = true; *hasResultType = true; break;
@@ -1918,30 +1925,6 @@ inline void SpvHasResultAndType(SpvOp opcode, bool *hasResult, bool *hasResultTy
     case SpvOpTerminateRayNV: *hasResult = false; *hasResultType = false; break;
     case SpvOpTraceNV: *hasResult = false; *hasResultType = false; break;
     case SpvOpTypeAccelerationStructureNV: *hasResult = true; *hasResultType = false; break;
-    case SpvOpTypeRayQueryProvisionalKHR: *hasResult = true; *hasResultType = false; break;
-    case SpvOpRayQueryInitializeKHR: *hasResult = false; *hasResultType = false; break;
-    case SpvOpRayQueryTerminateKHR: *hasResult = false; *hasResultType = false; break;
-    case SpvOpRayQueryGenerateIntersectionKHR: *hasResult = false; *hasResultType = false; break;
-    case SpvOpRayQueryConfirmIntersectionKHR: *hasResult = false; *hasResultType = false; break;
-    case SpvOpRayQueryProceedKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionTypeKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetRayTMinKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetRayFlagsKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionTKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionInstanceCustomIndexKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionInstanceIdKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionGeometryIndexKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionPrimitiveIndexKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionBarycentricsKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionFrontFaceKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionCandidateAABBOpaqueKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionObjectRayDirectionKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionObjectRayOriginKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetWorldRayDirectionKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetWorldRayOriginKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionObjectToWorldKHR: *hasResult = true; *hasResultType = true; break;
-    case SpvOpRayQueryGetIntersectionWorldToObjectKHR: *hasResult = true; *hasResultType = true; break;
     case SpvOpExecuteCallableNV: *hasResult = false; *hasResultType = false; break;
     case SpvOpTypeCooperativeMatrixNV: *hasResult = true; *hasResultType = false; break;
     case SpvOpCooperativeMatrixLoadNV: *hasResult = true; *hasResultType = true; break;
@@ -2096,6 +2079,23 @@ inline void SpvHasResultAndType(SpvOp opcode, bool *hasResult, bool *hasResultTy
     case SpvOpSubgroupAvcSicGetPackedSkcLumaCountThresholdINTEL: *hasResult = true; *hasResultType = true; break;
     case SpvOpSubgroupAvcSicGetPackedSkcLumaSumThresholdINTEL: *hasResult = true; *hasResultType = true; break;
     case SpvOpSubgroupAvcSicGetInterRawSadsINTEL: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetRayTMinKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetRayFlagsKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionTKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionInstanceCustomIndexKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionInstanceIdKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionGeometryIndexKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionPrimitiveIndexKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionBarycentricsKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionFrontFaceKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionCandidateAABBOpaqueKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionObjectRayDirectionKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionObjectRayOriginKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetWorldRayDirectionKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetWorldRayOriginKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionObjectToWorldKHR: *hasResult = true; *hasResultType = true; break;
+    case SpvOpRayQueryGetIntersectionWorldToObjectKHR: *hasResult = true; *hasResultType = true; break;
     }
 }
 #endif /* SPV_ENABLE_UTILITY_CODE */

3rdparty/spirv-cross/spirv.hpp

@@ -50,11 +50,11 @@ namespace spv {
 typedef unsigned int Id;
 
 #define SPV_VERSION 0x10500
-#define SPV_REVISION 1
+#define SPV_REVISION 3
 
 static const unsigned int MagicNumber = 0x07230203;
 static const unsigned int Version = 0x00010500;
-static const unsigned int Revision = 1;
+static const unsigned int Revision = 3;
 static const unsigned int OpCodeMask = 0xffff;
 static const unsigned int WordCountShift = 16;
 
@@ -1895,6 +1895,13 @@ inline void HasResultAndType(Op opcode, bool *hasResult, bool *hasResultType) {
     case OpSubgroupAnyKHR: *hasResult = true; *hasResultType = true; break;
     case OpSubgroupAllEqualKHR: *hasResult = true; *hasResultType = true; break;
     case OpSubgroupReadInvocationKHR: *hasResult = true; *hasResultType = true; break;
+    case OpTypeRayQueryProvisionalKHR: *hasResult = true; *hasResultType = false; break;
+    case OpRayQueryInitializeKHR: *hasResult = false; *hasResultType = false; break;
+    case OpRayQueryTerminateKHR: *hasResult = false; *hasResultType = false; break;
+    case OpRayQueryGenerateIntersectionKHR: *hasResult = false; *hasResultType = false; break;
+    case OpRayQueryConfirmIntersectionKHR: *hasResult = false; *hasResultType = false; break;
+    case OpRayQueryProceedKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionTypeKHR: *hasResult = true; *hasResultType = true; break;
     case OpGroupIAddNonUniformAMD: *hasResult = true; *hasResultType = true; break;
     case OpGroupFAddNonUniformAMD: *hasResult = true; *hasResultType = true; break;
     case OpGroupFMinNonUniformAMD: *hasResult = true; *hasResultType = true; break;
@@ -1914,30 +1921,6 @@ inline void HasResultAndType(Op opcode, bool *hasResult, bool *hasResultType) {
     case OpTerminateRayNV: *hasResult = false; *hasResultType = false; break;
     case OpTraceNV: *hasResult = false; *hasResultType = false; break;
     case OpTypeAccelerationStructureNV: *hasResult = true; *hasResultType = false; break;
-    case OpTypeRayQueryProvisionalKHR: *hasResult = true; *hasResultType = false; break;
-    case OpRayQueryInitializeKHR: *hasResult = false; *hasResultType = false; break;
-    case OpRayQueryTerminateKHR: *hasResult = false; *hasResultType = false; break;
-    case OpRayQueryGenerateIntersectionKHR: *hasResult = false; *hasResultType = false; break;
-    case OpRayQueryConfirmIntersectionKHR: *hasResult = false; *hasResultType = false; break;
-    case OpRayQueryProceedKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionTypeKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetRayTMinKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetRayFlagsKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionTKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionInstanceCustomIndexKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionInstanceIdKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionGeometryIndexKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionPrimitiveIndexKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionBarycentricsKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionFrontFaceKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionCandidateAABBOpaqueKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionObjectRayDirectionKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionObjectRayOriginKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetWorldRayDirectionKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetWorldRayOriginKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionObjectToWorldKHR: *hasResult = true; *hasResultType = true; break;
-    case OpRayQueryGetIntersectionWorldToObjectKHR: *hasResult = true; *hasResultType = true; break;
     case OpExecuteCallableNV: *hasResult = false; *hasResultType = false; break;
     case OpTypeCooperativeMatrixNV: *hasResult = true; *hasResultType = false; break;
     case OpCooperativeMatrixLoadNV: *hasResult = true; *hasResultType = true; break;
@@ -2092,6 +2075,23 @@ inline void HasResultAndType(Op opcode, bool *hasResult, bool *hasResultType) {
     case OpSubgroupAvcSicGetPackedSkcLumaCountThresholdINTEL: *hasResult = true; *hasResultType = true; break;
     case OpSubgroupAvcSicGetPackedSkcLumaSumThresholdINTEL: *hasResult = true; *hasResultType = true; break;
     case OpSubgroupAvcSicGetInterRawSadsINTEL: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetRayTMinKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetRayFlagsKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionTKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionInstanceCustomIndexKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionInstanceIdKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionGeometryIndexKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionPrimitiveIndexKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionBarycentricsKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionFrontFaceKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionCandidateAABBOpaqueKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionObjectRayDirectionKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionObjectRayOriginKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetWorldRayDirectionKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetWorldRayOriginKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionObjectToWorldKHR: *hasResult = true; *hasResultType = true; break;
+    case OpRayQueryGetIntersectionWorldToObjectKHR: *hasResult = true; *hasResultType = true; break;
     }
 }
 #endif /* SPV_ENABLE_UTILITY_CODE */

3rdparty/spirv-cross/spirv_common.hpp

@@ -563,6 +563,10 @@ struct SPIRType : IVariant
 
 	SmallVector<TypeID> member_types;
 
+	// If member order has been rewritten to handle certain scenarios with Offset,
+	// allow codegen to rewrite the index.
+	SmallVector<uint32_t> member_type_index_redirection;
+
 	struct ImageType
 	{
 		TypeID type;
@@ -776,7 +780,7 @@ struct SPIRBlock : IVariant
 		ComplexLoop
 	};
 
-	enum
+	enum : uint32_t
 	{
 		NoDominator = 0xffffffffu
 	};

3rdparty/spirv-cross/spirv_cross.cpp

@@ -4204,19 +4204,22 @@ Bitset Compiler::combined_decoration_for_member(const SPIRType &type, uint32_t i
 
 	if (type_meta)
 	{
-		auto &memb = type_meta->members;
-		if (index >= memb.size())
+		auto &members = type_meta->members;
+		if (index >= members.size())
 			return flags;
-		auto &dec = memb[index];
+		auto &dec = members[index];
 
-		// If our type is a struct, traverse all the members as well recursively.
 		flags.merge_or(dec.decoration_flags);
 
-		for (uint32_t i = 0; i < type.member_types.size(); i++)
+		auto &member_type = get<SPIRType>(type.member_types[index]);
+
+		// If our member type is a struct, traverse all the child members as well recursively.
+		auto &member_childs = member_type.member_types;
+		for (uint32_t i = 0; i < member_childs.size(); i++)
 		{
-			auto &memb_type = get<SPIRType>(type.member_types[i]);
-			if (!memb_type.pointer)
-				flags.merge_or(combined_decoration_for_member(memb_type, i));
+			auto &child_member_type = get<SPIRType>(member_childs[i]);
+			if (!child_member_type.pointer)
+				flags.merge_or(combined_decoration_for_member(member_type, i));
 		}
 	}
 

3rdparty/spirv-cross/spirv_glsl.cpp

@@ -9794,15 +9794,33 @@ void CompilerGLSL::emit_instruction(const Instruction &instruction)
 	}
 
 	case OpAtomicLoad:
-		flush_all_atomic_capable_variables();
-		// FIXME: Image?
-		// OpAtomicLoad seems to only be relevant for atomic counters.
+	{
+		// In plain GLSL, we have no atomic loads, so emulate this by fetch adding by 0 and hope compiler figures it out.
+		// Alternatively, we could rely on KHR_memory_model, but that's not very helpful for GL.
+		auto &type = expression_type(ops[2]);
 		forced_temporaries.insert(ops[1]);
-		GLSL_UFOP(atomicCounter);
+		bool atomic_image = check_atomic_image(ops[2]);
+		bool unsigned_type = (type.basetype == SPIRType::UInt) ||
+		                     (atomic_image && get<SPIRType>(type.image.type).basetype == SPIRType::UInt);
+		const char *op = atomic_image ? "imageAtomicAdd" : "atomicAdd";
+		const char *increment = unsigned_type ? "0u" : "0";
+		emit_op(ops[0], ops[1], join(op, "(", to_expression(ops[2]), ", ", increment, ")"), false);
+		flush_all_atomic_capable_variables();
 		break;
+	}
 
 	case OpAtomicStore:
-		SPIRV_CROSS_THROW("Unsupported opcode OpAtomicStore.");
+	{
+		// In plain GLSL, we have no atomic stores, so emulate this with an atomic exchange where we don't consume the result.
+		// Alternatively, we could rely on KHR_memory_model, but that's not very helpful for GL.
+		uint32_t ptr = ops[0];
+		// Ignore semantics for now, probably only relevant to CL.
+		uint32_t val = ops[3];
+		const char *op = check_atomic_image(ptr) ? "imageAtomicExchange" : "atomicExchange";
+		statement(op, "(", to_expression(ptr), ", ", to_expression(val), ");");
+		flush_all_atomic_capable_variables();
+		break;
+	}
 
 	case OpAtomicIIncrement:
 	case OpAtomicIDecrement:

3rdparty/spirv-cross/spirv_hlsl.cpp

@@ -4087,9 +4087,11 @@ void CompilerHLSL::emit_atomic(const uint32_t *ops, uint32_t length, spv::Op op)
 	const char *atomic_op = nullptr;
 
 	string value_expr;
-	if (op != OpAtomicIDecrement && op != OpAtomicIIncrement)
+	if (op != OpAtomicIDecrement && op != OpAtomicIIncrement && op != OpAtomicLoad && op != OpAtomicStore)
 		value_expr = to_expression(ops[op == OpAtomicCompareExchange ? 6 : 5]);
 
+	bool is_atomic_store = false;
+
 	switch (op)
 	{
 	case OpAtomicIIncrement:
@@ -4102,6 +4104,11 @@ void CompilerHLSL::emit_atomic(const uint32_t *ops, uint32_t length, spv::Op op)
 		value_expr = "-1";
 		break;
 
+	case OpAtomicLoad:
+		atomic_op = "InterlockedAdd";
+		value_expr = "0";
+		break;
+
 	case OpAtomicISub:
 		atomic_op = "InterlockedAdd";
 		value_expr = join("-", enclose_expression(value_expr));
@@ -4137,6 +4144,11 @@ void CompilerHLSL::emit_atomic(const uint32_t *ops, uint32_t length, spv::Op op)
 		atomic_op = "InterlockedExchange";
 		break;
 
+	case OpAtomicStore:
+		atomic_op = "InterlockedExchange";
+		is_atomic_store = true;
+		break;
+
 	case OpAtomicCompareExchange:
 		if (length < 8)
 			SPIRV_CROSS_THROW("Not enough data for opcode.");
@@ -4148,31 +4160,57 @@ void CompilerHLSL::emit_atomic(const uint32_t *ops, uint32_t length, spv::Op op)
 		SPIRV_CROSS_THROW("Unknown atomic opcode.");
 	}
 
-	uint32_t result_type = ops[0];
-	uint32_t id = ops[1];
-	forced_temporaries.insert(ops[1]);
+	if (is_atomic_store)
+	{
+		auto &data_type = expression_type(ops[0]);
+		auto *chain = maybe_get<SPIRAccessChain>(ops[0]);
 
-	auto &type = get<SPIRType>(result_type);
-	statement(variable_decl(type, to_name(id)), ";");
+		auto &tmp_id = extra_sub_expressions[ops[0]];
+		if (!tmp_id)
+		{
+			tmp_id = ir.increase_bound_by(1);
+			emit_uninitialized_temporary_expression(get_pointee_type(data_type).self, tmp_id);
+		}
 
-	auto &data_type = expression_type(ops[2]);
-	auto *chain = maybe_get<SPIRAccessChain>(ops[2]);
-	SPIRType::BaseType expr_type;
-	if (data_type.storage == StorageClassImage || !chain)
-	{
-		statement(atomic_op, "(", to_expression(ops[2]), ", ", value_expr, ", ", to_name(id), ");");
-		expr_type = data_type.basetype;
+		if (data_type.storage == StorageClassImage || !chain)
+		{
+			statement(atomic_op, "(", to_expression(ops[0]), ", ", to_expression(ops[3]), ", ", to_expression(tmp_id), ");");
+		}
+		else
+		{
+			// RWByteAddress buffer is always uint in its underlying type.
+			statement(chain->base, ".", atomic_op, "(", chain->dynamic_index, chain->static_index, ", ", to_expression(ops[3]),
+			          ", ", to_expression(tmp_id), ");");
+		}
 	}
 	else
 	{
-		// RWByteAddress buffer is always uint in its underlying type.
-		expr_type = SPIRType::UInt;
-		statement(chain->base, ".", atomic_op, "(", chain->dynamic_index, chain->static_index, ", ", value_expr, ", ",
-		          to_name(id), ");");
-	}
+		uint32_t result_type = ops[0];
+		uint32_t id = ops[1];
+		forced_temporaries.insert(ops[1]);
+
+		auto &type = get<SPIRType>(result_type);
+		statement(variable_decl(type, to_name(id)), ";");
 
-	auto expr = bitcast_expression(type, expr_type, to_name(id));
-	set<SPIRExpression>(id, expr, result_type, true);
+		auto &data_type = expression_type(ops[2]);
+		auto *chain = maybe_get<SPIRAccessChain>(ops[2]);
+		SPIRType::BaseType expr_type;
+		if (data_type.storage == StorageClassImage || !chain)
+		{
+			statement(atomic_op, "(", to_expression(ops[2]), ", ", value_expr, ", ", to_name(id), ");");
+			expr_type = data_type.basetype;
+		}
+		else
+		{
+			// RWByteAddress buffer is always uint in its underlying type.
+			expr_type = SPIRType::UInt;
+			statement(chain->base, ".", atomic_op, "(", chain->dynamic_index, chain->static_index, ", ", value_expr,
+			          ", ", to_name(id), ");");
+		}
+
+		auto expr = bitcast_expression(type, expr_type, to_name(id));
+		set<SPIRExpression>(id, expr, result_type, true);
+	}
 	flush_all_atomic_capable_variables();
 }
 
@@ -4967,6 +5005,8 @@ void CompilerHLSL::emit_instruction(const Instruction &instruction)
 	case OpAtomicIAdd:
 	case OpAtomicIIncrement:
 	case OpAtomicIDecrement:
+	case OpAtomicLoad:
+	case OpAtomicStore:
 	{
 		emit_atomic(ops, instruction.length, opcode);
 		break;

3rdparty/spirv-cross/spirv_msl.cpp

@@ -7038,7 +7038,11 @@ void CompilerMSL::emit_atomic_func_op(uint32_t result_type, uint32_t result_id,
 			exp += string(", ") + get_memory_order(mem_order_2);
 
 		exp += ")";
-		emit_op(result_type, result_id, exp, false);
+
+		if (strcmp(op, "atomic_store_explicit") != 0)
+			emit_op(result_type, result_id, exp, false);
+		else
+			statement(exp, ";");
 	}
 
 	flush_all_atomic_capable_variables();
@@ -9185,7 +9189,14 @@ string CompilerMSL::get_type_address_space(const SPIRType &type, uint32_t id, bo
 				addr_space = "constant";
 		}
 		else if (!argument)
+		{
 			addr_space = "constant";
+		}
+		else if (type_is_msl_framebuffer_fetch(type))
+		{
+			// Subpass inputs are passed around by value.
+			addr_space = "";
+		}
 		break;
 
 	case StorageClassFunction:
@@ -9626,8 +9637,7 @@ void CompilerMSL::entry_point_args_discrete_descriptors(string &ep_args)
 
 			// Use Metal's native frame-buffer fetch API for subpass inputs.
 			const auto &basetype = get<SPIRType>(var.basetype);
-			if (basetype.image.dim != DimSubpassData || !msl_options.is_ios() ||
-			    !msl_options.ios_use_framebuffer_fetch_subpasses)
+			if (!type_is_msl_framebuffer_fetch(basetype))
 			{
 				ep_args += image_type_glsl(type, var_id) + " " + r.name;
 				if (r.plane > 0)
@@ -9639,7 +9649,7 @@ void CompilerMSL::entry_point_args_discrete_descriptors(string &ep_args)
 			}
 			else
 			{
-				ep_args += image_type_glsl(type, var_id) + "4 " + r.name;
+				ep_args += image_type_glsl(type, var_id) + " " + r.name;
 				ep_args += " [[color(" + convert_to_string(r.index) + ")]]";
 			}
 
@@ -10125,6 +10135,12 @@ uint32_t CompilerMSL::get_metal_resource_index(SPIRVariable &var, SPIRType::Base
 	return resource_index;
 }
 
+bool CompilerMSL::type_is_msl_framebuffer_fetch(const SPIRType &type) const
+{
+	return type.basetype == SPIRType::Image && type.image.dim == DimSubpassData &&
+	       msl_options.is_ios() && msl_options.ios_use_framebuffer_fetch_subpasses;
+}
+
 string CompilerMSL::argument_decl(const SPIRFunction::Parameter &arg)
 {
 	auto &var = get<SPIRVariable>(arg.id);
@@ -10140,6 +10156,9 @@ string CompilerMSL::argument_decl(const SPIRFunction::Parameter &arg)
 		name_id = var.basevariable;
 
 	bool constref = !arg.alias_global_variable && is_pointer && arg.write_count == 0;
+	// Framebuffer fetch is plain value, const looks out of place, but it is not wrong.
+	if (type_is_msl_framebuffer_fetch(type))
+		constref = false;
 
 	bool type_is_image = type.basetype == SPIRType::Image || type.basetype == SPIRType::SampledImage ||
 	                     type.basetype == SPIRType::Sampler;
@@ -10638,6 +10657,9 @@ void CompilerMSL::replace_illegal_names()
 
 string CompilerMSL::to_member_reference(uint32_t base, const SPIRType &type, uint32_t index, bool ptr_chain)
 {
+	if (index < uint32_t(type.member_type_index_redirection.size()))
+		index = type.member_type_index_redirection[index];
+
 	auto *var = maybe_get<SPIRVariable>(base);
 	// If this is a buffer array, we have to dereference the buffer pointers.
 	// Otherwise, if this is a pointer expression, dereference it.
@@ -10983,10 +11005,11 @@ string CompilerMSL::image_type_glsl(const SPIRType &type, uint32_t id)
 			}
 
 			// Use Metal's native frame-buffer fetch API for subpass inputs.
-			if (img_type.dim == DimSubpassData && msl_options.is_ios() &&
-			    msl_options.ios_use_framebuffer_fetch_subpasses)
+			if (type_is_msl_framebuffer_fetch(type))
 			{
-				return type_to_glsl(get<SPIRType>(img_type.type));
+				auto img_type_4 = get<SPIRType>(img_type.type);
+				img_type_4.vecsize = 4;
+				return type_to_glsl(img_type_4);
 			}
 			if (img_type.ms && img_type.arrayed)
 			{
@@ -12437,9 +12460,28 @@ void CompilerMSL::MemberSorter::sort()
 	// the members should be reordered, based on builtin and sorting aspect meta info.
 	size_t mbr_cnt = type.member_types.size();
 	SmallVector<uint32_t> mbr_idxs(mbr_cnt);
-	iota(mbr_idxs.begin(), mbr_idxs.end(), 0); // Fill with consecutive indices
+	std::iota(mbr_idxs.begin(), mbr_idxs.end(), 0); // Fill with consecutive indices
 	std::stable_sort(mbr_idxs.begin(), mbr_idxs.end(), *this); // Sort member indices based on sorting aspect
 
+	bool sort_is_identity = true;
+	for (uint32_t mbr_idx = 0; mbr_idx < mbr_cnt; mbr_idx++)
+	{
+		if (mbr_idx != mbr_idxs[mbr_idx])
+		{
+			sort_is_identity = false;
+			break;
+		}
+	}
+
+	if (sort_is_identity)
+		return;
+
+	if (meta.members.size() < type.member_types.size())
+	{
+		// This should never trigger in normal circumstances, but to be safe.
+		meta.members.resize(type.member_types.size());
+	}
+
 	// Move type and meta member info to the order defined by the sorted member indices.
 	// This is done by creating temporary copies of both member types and meta, and then
 	// copying back to the original content at the sorted indices.
@@ -12450,6 +12492,13 @@ void CompilerMSL::MemberSorter::sort()
 		type.member_types[mbr_idx] = mbr_types_cpy[mbr_idxs[mbr_idx]];
 		meta.members[mbr_idx] = mbr_meta_cpy[mbr_idxs[mbr_idx]];
 	}
+
+	if (sort_aspect == SortAspect::Offset)
+	{
+		// If we're sorting by Offset, this might affect user code which accesses a buffer block.
+		// We will need to redirect member indices from one index to sorted index.
+		type.member_type_index_redirection = std::move(mbr_idxs);
+	}
 }
 
 // Sort first by builtin status (put builtins at end), then by the sorting aspect.

3rdparty/spirv-cross/spirv_msl.hpp

@@ -324,12 +324,12 @@ public:
 		// can be read in subsequent stages.
 		bool enable_clip_distance_user_varying = true;
 
-		bool is_ios()
+		bool is_ios() const
 		{
 			return platform == iOS;
 		}
 
-		bool is_macos()
+		bool is_macos() const
 		{
 			return platform == macOS;
 		}
@@ -898,6 +898,8 @@ protected:
 
 	void activate_argument_buffer_resources();
 
+	bool type_is_msl_framebuffer_fetch(const SPIRType &type) const;
+
 	// OpcodeHandler that handles several MSL preprocessing operations.
 	struct OpCodePreprocessor : OpcodeHandler
 	{