implement DS_OR_B32

src/shader_recompiler/backend/spirv/emit_spirv_atomic.cpp

@@ -60,6 +60,10 @@ Id EmitSharedAtomicSMin32(EmitContext& ctx, Id offset, Id value) {
     return SharedAtomicU32(ctx, offset, value, &Sirit::Module::OpAtomicSMin);
 }
 
+Id EmitSharedAtomicOr32(EmitContext& ctx, Id offset, Id value) {
+    return SharedAtomicU32(ctx, offset, value, &Sirit::Module::OpAtomicOr);
+}
+
 Id EmitBufferAtomicIAdd32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id address, Id value) {
     return BufferAtomicU32(ctx, inst, handle, address, value, &Sirit::Module::OpAtomicIAdd);
 }

src/shader_recompiler/backend/spirv/emit_spirv_instructions.h

@@ -112,6 +112,7 @@ Id EmitSharedAtomicUMax32(EmitContext& ctx, Id offset, Id value);
 Id EmitSharedAtomicSMax32(EmitContext& ctx, Id offset, Id value);
 Id EmitSharedAtomicUMin32(EmitContext& ctx, Id offset, Id value);
 Id EmitSharedAtomicSMin32(EmitContext& ctx, Id offset, Id value);
+Id EmitSharedAtomicOr32(EmitContext& ctx, Id offset, Id value);
 Id EmitCompositeConstructU32x2(EmitContext& ctx, Id e1, Id e2);
 Id EmitCompositeConstructU32x3(EmitContext& ctx, Id e1, Id e2, Id e3);
 Id EmitCompositeConstructU32x4(EmitContext& ctx, Id e1, Id e2, Id e3, Id e4);

src/shader_recompiler/frontend/translate/data_share.cpp

@@ -50,6 +50,8 @@ void Translator::EmitDataShare(const GcnInst& inst) {
         return DS_READ(64, false, false, false, inst);
     case Opcode::DS_READ2_B64:
         return DS_READ(64, false, true, false, inst);
+    case Opcode::DS_OR_B32:
+        return DS_OR_B32(inst, false);
     default:
         LogMissingOpcode(inst);
     }
@@ -119,6 +121,18 @@ void Translator::DS_MAX_U32(const GcnInst& inst, bool is_signed, bool rtn) {
     }
 }
 
+void Translator::DS_OR_B32(const GcnInst& inst, bool rtn) {
+    const IR::U32 addr{GetSrc(inst.src[0])};
+    const IR::U32 data{GetSrc(inst.src[1])};
+    const IR::U32 offset =
+        ir.Imm32((u32(inst.control.ds.offset1) << 8u) + u32(inst.control.ds.offset0));
+    const IR::U32 addr_offset = ir.IAdd(addr, offset);
+    const IR::Value original_val = ir.SharedAtomicOr(addr_offset, data);
+    if (rtn) {
+        SetDst(inst.dst[0], IR::U32{original_val});
+    }
+}
+
 void Translator::DS_WRITE(int bit_size, bool is_signed, bool is_pair, bool stride64,
                           const GcnInst& inst) {
     const IR::U32 addr{ir.GetVectorReg(IR::VectorReg(inst.src[0].code))};

src/shader_recompiler/frontend/translate/translate.h

@@ -250,6 +250,7 @@ public:
     void DS_MAX_U32(const GcnInst& inst, bool is_signed, bool rtn);
     void DS_WRITE(int bit_size, bool is_signed, bool is_pair, bool stride64, const GcnInst& inst);
     void DS_SWIZZLE_B32(const GcnInst& inst);
+    void DS_OR_B32(const GcnInst& inst, bool rtn);
     void DS_READ(int bit_size, bool is_signed, bool is_pair, bool stride64, const GcnInst& inst);
     void DS_APPEND(const GcnInst& inst);
     void DS_CONSUME(const GcnInst& inst);

src/shader_recompiler/ir/ir_emitter.cpp

@@ -326,6 +326,10 @@ U32 IREmitter::SharedAtomicIMax(const U32& address, const U32& data, bool is_sig
                      : Inst<U32>(Opcode::SharedAtomicUMax32, address, data);
 }
 
+U32 IREmitter::SharedAtomicOr(const U32& address, const U32& data) {
+    return Inst<U32>(Opcode::SharedAtomicOr32, address, data);
+}
+
 U32 IREmitter::ReadConst(const Value& base, const U32& offset) {
     return Inst<U32>(Opcode::ReadConst, base, offset);
 }

src/shader_recompiler/ir/ir_emitter.h

@@ -90,6 +90,7 @@ public:
     [[nodiscard]] U32F32 SharedAtomicIAdd(const U32& address, const U32F32& data);
     [[nodiscard]] U32 SharedAtomicIMin(const U32& address, const U32& data, bool is_signed);
     [[nodiscard]] U32 SharedAtomicIMax(const U32& address, const U32& data, bool is_signed);
+    [[nodiscard]] U32 SharedAtomicOr(const U32& address, const U32& data);
 
     [[nodiscard]] U32 ReadConst(const Value& base, const U32& offset);
     [[nodiscard]] U32 ReadConstBuffer(const Value& handle, const U32& index);

src/shader_recompiler/ir/opcodes.inc

@@ -43,6 +43,7 @@ OPCODE(SharedAtomicSMin32,                                  U32,            U32,
 OPCODE(SharedAtomicUMin32,                                  U32,            U32,            U32,                                                            )
 OPCODE(SharedAtomicSMax32,                                  U32,            U32,            U32,                                                            )
 OPCODE(SharedAtomicUMax32,                                  U32,            U32,            U32,                                                            )
+OPCODE(SharedAtomicOr32,                                    U32,            U32,            U32,                                                            )
 
 // Context getters/setters
 OPCODE(GetUserData,                                         U32,            ScalarReg,                                                                      )