Merge pull request #3347 from ReinUsesLisp/local-mem

 shader/memory: Implement LDL.S16, LDS.S16, STL.S16 and STS.S16

src/video_core/shader/decode/memory.cpp

@@ -24,6 +24,7 @@ using Tegra::Shader::GlobalAtomicType;
 using Tegra::Shader::Instruction;
 using Tegra::Shader::OpCode;
 using Tegra::Shader::Register;
+using Tegra::Shader::StoreType;
 
 namespace {
 
@@ -63,6 +64,27 @@ u32 GetMemorySize(Tegra::Shader::UniformType uniform_type) {
     }
 }
 
+Node ExtractUnaligned(Node value, Node address, u32 mask, u32 size) {
+    Node offset = Operation(OperationCode::UBitwiseAnd, address, Immediate(mask));
+    offset = Operation(OperationCode::ULogicalShiftLeft, std::move(offset), Immediate(3));
+    return Operation(OperationCode::UBitfieldExtract, std::move(value), std::move(offset),
+                     Immediate(size));
+}
+
+Node InsertUnaligned(Node dest, Node value, Node address, u32 mask, u32 size) {
+    Node offset = Operation(OperationCode::UBitwiseAnd, std::move(address), Immediate(mask));
+    offset = Operation(OperationCode::ULogicalShiftLeft, std::move(offset), Immediate(3));
+    return Operation(OperationCode::UBitfieldInsert, std::move(dest), std::move(value),
+                     std::move(offset), Immediate(size));
+}
+
+Node Sign16Extend(Node value) {
+    Node sign = Operation(OperationCode::UBitwiseAnd, value, Immediate(1U << 15));
+    Node is_sign = Operation(OperationCode::LogicalUEqual, std::move(sign), Immediate(1U << 15));
+    Node extend = Operation(OperationCode::Select, is_sign, Immediate(0xFFFF0000), Immediate(0));
+    return Operation(OperationCode::UBitwiseOr, std::move(value), std::move(extend));
+}
+
 } // Anonymous namespace
 
 u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) {
@@ -138,26 +160,31 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) {
         LOG_DEBUG(HW_GPU, "LD_L cache management mode: {}", static_cast<u64>(instr.ld_l.unknown));
         [[fallthrough]];
     case OpCode::Id::LD_S: {
-        const auto GetMemory = [&](s32 offset) {
+        const auto GetAddress = [&](s32 offset) {
             ASSERT(offset % 4 == 0);
             const Node immediate_offset = Immediate(static_cast<s32>(instr.smem_imm) + offset);
-            const Node address = Operation(OperationCode::IAdd, NO_PRECISE, GetRegister(instr.gpr8),
-                                           immediate_offset);
-            return opcode->get().GetId() == OpCode::Id::LD_S ? GetSharedMemory(address)
-                                                             : GetLocalMemory(address);
+            return Operation(OperationCode::IAdd, GetRegister(instr.gpr8), immediate_offset);
+        };
+        const auto GetMemory = [&](s32 offset) {
+            return opcode->get().GetId() == OpCode::Id::LD_S ? GetSharedMemory(GetAddress(offset))
+                                                             : GetLocalMemory(GetAddress(offset));
         };
 
         switch (instr.ldst_sl.type.Value()) {
-        case Tegra::Shader::StoreType::Bits32:
-        case Tegra::Shader::StoreType::Bits64:
-        case Tegra::Shader::StoreType::Bits128: {
-            const u32 count = [&]() {
+        case StoreType::Signed16:
+            SetRegister(bb, instr.gpr0,
+                        Sign16Extend(ExtractUnaligned(GetMemory(0), GetAddress(0), 0b10, 16)));
+            break;
+        case StoreType::Bits32:
+        case StoreType::Bits64:
+        case StoreType::Bits128: {
+            const u32 count = [&] {
                 switch (instr.ldst_sl.type.Value()) {
-                case Tegra::Shader::StoreType::Bits32:
+                case StoreType::Bits32:
                     return 1;
-                case Tegra::Shader::StoreType::Bits64:
+                case StoreType::Bits64:
                     return 2;
-                case Tegra::Shader::StoreType::Bits128:
+                case StoreType::Bits128:
                     return 4;
                 default:
                     UNREACHABLE();
@@ -214,12 +241,7 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) {
             // To handle unaligned loads get the bytes used to dereference global memory and extract
             // those bytes from the loaded u32.
             if (IsUnaligned(type)) {
-                Node mask = Immediate(GetUnalignedMask(type));
-                Node offset = Operation(OperationCode::UBitwiseAnd, real_address, std::move(mask));
-                offset = Operation(OperationCode::ULogicalShiftLeft, offset, Immediate(3));
-
-                gmem = Operation(OperationCode::UBitfieldExtract, std::move(gmem),
-                                 std::move(offset), Immediate(size));
+                gmem = ExtractUnaligned(gmem, real_address, GetUnalignedMask(type), size);
             }
 
             SetTemporary(bb, i, gmem);
@@ -271,21 +293,28 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) {
             return Operation(OperationCode::IAdd, NO_PRECISE, GetRegister(instr.gpr8), immediate);
         };
 
-        const auto set_memory = opcode->get().GetId() == OpCode::Id::ST_L
-                                    ? &ShaderIR::SetLocalMemory
-                                    : &ShaderIR::SetSharedMemory;
+        const bool is_local = opcode->get().GetId() == OpCode::Id::ST_L;
+        const auto set_memory = is_local ? &ShaderIR::SetLocalMemory : &ShaderIR::SetSharedMemory;
+        const auto get_memory = is_local ? &ShaderIR::GetLocalMemory : &ShaderIR::GetSharedMemory;
 
         switch (instr.ldst_sl.type.Value()) {
-        case Tegra::Shader::StoreType::Bits128:
+        case StoreType::Bits128:
             (this->*set_memory)(bb, GetAddress(12), GetRegister(instr.gpr0.Value() + 3));
             (this->*set_memory)(bb, GetAddress(8), GetRegister(instr.gpr0.Value() + 2));
             [[fallthrough]];
-        case Tegra::Shader::StoreType::Bits64:
+        case StoreType::Bits64:
             (this->*set_memory)(bb, GetAddress(4), GetRegister(instr.gpr0.Value() + 1));
             [[fallthrough]];
-        case Tegra::Shader::StoreType::Bits32:
+        case StoreType::Bits32:
             (this->*set_memory)(bb, GetAddress(0), GetRegister(instr.gpr0));
             break;
+        case StoreType::Signed16: {
+            Node address = GetAddress(0);
+            Node memory = (this->*get_memory)(address);
+            (this->*set_memory)(
+                bb, address, InsertUnaligned(memory, GetRegister(instr.gpr0), address, 0b10, 16));
+            break;
+        }
         default:
             UNIMPLEMENTED_MSG("{} unhandled type: {}", opcode->get().GetName(),
                               static_cast<u32>(instr.ldst_sl.type.Value()));
@@ -325,12 +354,8 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) {
             Node value = GetRegister(instr.gpr0.Value() + i);
 
             if (IsUnaligned(type)) {
-                Node mask = Immediate(GetUnalignedMask(type));
-                Node offset = Operation(OperationCode::UBitwiseAnd, real_address, std::move(mask));
-                offset = Operation(OperationCode::ULogicalShiftLeft, offset, Immediate(3));
-
-                value = Operation(OperationCode::UBitfieldInsert, gmem, std::move(value), offset,
-                                  Immediate(size));
+                const u32 mask = GetUnalignedMask(type);
+                value = InsertUnaligned(gmem, std::move(value), real_address, mask, size);
             }
 
             bb.push_back(Operation(OperationCode::Assign, gmem, value));