Merge pull request #486 from lioncash/barrier

A32: Implement barrier instructions introduced in ARMv7

src/CMakeLists.txt

@@ -95,6 +95,7 @@ add_library(dynarmic
     frontend/A32/translate/translate.cpp
     frontend/A32/translate/translate.h
     frontend/A32/translate/translate_arm.cpp
+    frontend/A32/translate/translate_arm/barrier.cpp
     frontend/A32/translate/translate_arm/branch.cpp
     frontend/A32/translate/translate_arm/coprocessor.cpp
     frontend/A32/translate/translate_arm/data_processing.cpp

src/backend/x64/a32_emit_x64.cpp

@@ -615,6 +615,23 @@ void A32EmitX64::EmitA32SetGEFlagsCompressed(A32EmitContext& ctx, IR::Inst* inst
     }
 }
 
+void A32EmitX64::EmitA32DataSynchronizationBarrier(A32EmitContext&, IR::Inst*) {
+    code.mfence();
+}
+
+void A32EmitX64::EmitA32DataMemoryBarrier(A32EmitContext&, IR::Inst*) {
+    code.lfence();
+}
+
+void A32EmitX64::EmitA32InstructionSynchronizationBarrier(A32EmitContext& ctx, IR::Inst*) {
+    ctx.reg_alloc.HostCall(nullptr);
+
+    code.mov(code.ABI_PARAM1, reinterpret_cast<u64>(jit_interface));
+    code.CallFunction(static_cast<void(*)(A32::Jit*)>([](A32::Jit* jit) {
+        jit->ClearCache();
+    }));
+}
+
 void A32EmitX64::EmitA32BXWritePC(A32EmitContext& ctx, IR::Inst* inst) {
     auto args = ctx.reg_alloc.GetArgumentInfo(inst);
     auto& arg = args[0];

src/frontend/A32/decoder/arm.inc

@@ -1,3 +1,8 @@
+// Barrier instructions
+INST(arm_DMB,           "DMB",                 "1111010101111111111100000101oooo") // v7
+INST(arm_DSB,           "DSB",                 "1111010101111111111100000100oooo") // v7
+INST(arm_ISB,           "ISB",                 "1111010101111111111100000110oooo") // v7
+
 // Branch instructions
 INST(arm_BLX_imm,       "BLX (imm)",           "1111101hvvvvvvvvvvvvvvvvvvvvvvvv") // v5
 INST(arm_BLX_reg,       "BLX (reg)",           "cccc000100101111111111110011mmmm") // v5

src/frontend/A32/disassembler/disassembler_arm.cpp

@@ -76,6 +76,29 @@ public:
         return "<internal error>";
     }
 
+    static const char* BarrierOptionStr(Imm4 option) {
+        switch (option) {
+        case 0b0010:
+            return " oshst";
+        case 0b0011:
+            return " osh";
+        case 0b0110:
+            return " nshst";
+        case 0b0111:
+            return " nsh";
+        case 0b1010:
+            return " ishst";
+        case 0b1011:
+            return " ish";
+        case 0b1110:
+            return " st";
+        case 0b1111: // SY can be omitted.
+            return "";
+        default:
+            return " unknown";
+        }
+    }
+
     std::string FPRegStr(bool dp_operation, size_t base, bool bit) {
         size_t reg_num;
         if (dp_operation) {
@@ -100,6 +123,17 @@ public:
         return cond == Cond::NV ? "2" : CondToString(cond);
     }
 
+    // Barrier instructions
+    std::string arm_DMB(Imm4 option) {
+        return fmt::format("dmb{}", BarrierOptionStr(option));
+    }
+    std::string arm_DSB(Imm4 option) {
+        return fmt::format("dsb{}", BarrierOptionStr(option));
+    }
+    std::string arm_ISB([[maybe_unused]] Imm4 option) {
+        return "isb";
+    }
+
     // Branch instructions
     std::string arm_B(Cond cond, Imm24 imm24) {
         s32 offset = Common::SignExtend<26, s32>(imm24 << 2) + 8;

src/frontend/A32/ir_emitter.cpp

@@ -142,6 +142,18 @@ void IREmitter::SetGEFlagsCompressed(const IR::U32& value) {
     Inst(Opcode::A32SetGEFlagsCompressed, value);
 }
 
+void IREmitter::DataSynchronizationBarrier() {
+    Inst(Opcode::A32DataSynchronizationBarrier);
+}
+
+void IREmitter::DataMemoryBarrier() {
+    Inst(Opcode::A32DataMemoryBarrier);
+}
+
+void IREmitter::InstructionSynchronizationBarrier() {
+    Inst(Opcode::A32InstructionSynchronizationBarrier);
+}
+
 IR::U32 IREmitter::GetFpscr() {
     return Inst<IR::U32>(Opcode::A32GetFpscr);
 }

src/frontend/A32/ir_emitter.h

@@ -61,6 +61,10 @@ public:
     void SetGEFlags(const IR::U32& value);
     void SetGEFlagsCompressed(const IR::U32& value);
 
+    void DataSynchronizationBarrier();
+    void DataMemoryBarrier();
+    void InstructionSynchronizationBarrier();
+
     IR::U32 GetFpscr();
     void SetFpscr(const IR::U32& new_fpscr);
     IR::U32 GetFpscrNZCV();

src/frontend/A32/translate/translate_arm/barrier.cpp

@@ -0,0 +1,28 @@
+/* This file is part of the dynarmic project.
+ * Copyright (c) 2019 MerryMage
+ * This software may be used and distributed according to the terms of the GNU
+ * General Public License version 2 or any later version.
+ */
+
+#include "translate_arm.h"
+
+namespace Dynarmic::A32 {
+
+bool ArmTranslatorVisitor::arm_DMB([[maybe_unused]] Imm4 option) {
+    ir.DataMemoryBarrier();
+    return true;
+}
+
+bool ArmTranslatorVisitor::arm_DSB([[maybe_unused]] Imm4 option) {
+    ir.DataSynchronizationBarrier();
+    return true;
+}
+
+bool ArmTranslatorVisitor::arm_ISB([[maybe_unused]] Imm4 option) {
+    ir.InstructionSynchronizationBarrier();
+    ir.SetRegister(Reg::PC, ir.Imm32(ir.current_location.PC() + 4));
+    ir.SetTerm(IR::Term::ReturnToDispatch{});
+    return false;
+}
+
+} // namespace Dynarmic::A32

src/frontend/A32/translate/translate_arm/translate_arm.h

@@ -64,6 +64,11 @@ struct ArmTranslatorVisitor final {
     template <typename FnT> bool EmitVfpVectorOperation(bool sz, ExtReg d, ExtReg n, ExtReg m, const FnT& fn);
     template <typename FnT> bool EmitVfpVectorOperation(bool sz, ExtReg d, ExtReg m, const FnT& fn);
 
+    // Barrier instructions
+    bool arm_DMB(Imm4 option);
+    bool arm_DSB(Imm4 option);
+    bool arm_ISB(Imm4 option);
+
     // Branch instructions
     bool arm_B(Cond cond, Imm24 imm24);
     bool arm_BL(Cond cond, Imm24 imm24);

src/frontend/ir/microinstruction.cpp

@@ -45,6 +45,21 @@ bool Inst::IsShift() const {
            IsLogicalShift();
 }
 
+bool Inst::IsBarrier() const {
+    switch (op) {
+    case Opcode::A32DataMemoryBarrier:
+    case Opcode::A32DataSynchronizationBarrier:
+    case Opcode::A32InstructionSynchronizationBarrier:
+    case Opcode::A64DataMemoryBarrier:
+    case Opcode::A64DataSynchronizationBarrier:
+    case Opcode::A64InstructionSynchronizationBarrier:
+        return true;
+
+    default:
+        return false;
+    }
+}
+
 bool Inst::IsSharedMemoryRead() const {
     switch (op) {
     case Opcode::A32ReadMemory8:
@@ -466,9 +481,7 @@ bool Inst::MayHaveSideEffects() const {
     return op == Opcode::PushRSB                        ||
            op == Opcode::A64SetCheckBit                 ||
            op == Opcode::A64DataCacheOperationRaised    ||
-           op == Opcode::A64DataSynchronizationBarrier        ||
+           IsBarrier()                                  ||
-           op == Opcode::A64DataMemoryBarrier                 ||
-           op == Opcode::A64InstructionSynchronizationBarrier ||
            CausesCPUException()                         ||
            WritesToCoreRegister()                       ||
            WritesToSystemRegister()                     ||

src/frontend/ir/microinstruction.h

@@ -36,6 +36,9 @@ public:
     /// Determines whether or not this instruction performs any kind of shift.
     bool IsShift() const;
 
+    /// Determines whether or not this instruction is a form of barrier.
+    bool IsBarrier() const;
+
     /// Determines whether or not this instruction performs a shared memory read.
     bool IsSharedMemoryRead() const;
     /// Determines whether or not this instruction performs a shared memory write.

src/frontend/ir/opcodes.inc

@@ -30,6 +30,9 @@ A32OPC(SetGEFlagsCompressed,                                Void,           U32
 A32OPC(BXWritePC,                                           Void,           U32                                                             )
 A32OPC(CallSupervisor,                                      Void,           U32                                                             )
 A32OPC(ExceptionRaised,                                     Void,           U32,            U64                                             )
+A32OPC(DataSynchronizationBarrier,                          Void,                                                                           )
+A32OPC(DataMemoryBarrier,                                   Void,                                                                           )
+A32OPC(InstructionSynchronizationBarrier,                   Void,                                                                           )
 A32OPC(GetFpscr,                                            U32,                                                                            )
 A32OPC(SetFpscr,                                            Void,           U32,                                                            )
 A32OPC(GetFpscrNZCV,                                        U32,                                                                            )