Merge 56d8d65e22 into 8ee64a84c7

Source/Core/Core/PowerPC/BreakPoints.cpp

@@ -346,8 +346,12 @@ void MemChecks::Update()
 {
   const Core::CPUThreadGuard guard(m_system);
 
-  // Clear the JIT cache so it can switch the watchpoint-compatible mode.
+  const bool registers_changed = UpdateRegistersUsedInConditions();
-  if (m_mem_breakpoints_set != HasAny())
+
+  // If we've added a first memcheck, clear the JIT cache so it can switch to watchpoint-compatible
+  // code. Or, if we've added a memcheck whose condition wants to read from a new register,
+  // clear the JIT cache to make the slow memory access code flush that register.
+  if (registers_changed || m_mem_breakpoints_set != HasAny())
   {
     m_system.GetJitInterface().ClearCache(guard);
     m_mem_breakpoints_set = HasAny();
@@ -356,6 +360,27 @@ void MemChecks::Update()
   m_system.GetMMU().DBATUpdated();
 }
 
+bool MemChecks::UpdateRegistersUsedInConditions()
+{
+  BitSet32 gprs_used, fprs_used;
+  for (TMemCheck& mem_check : m_mem_checks)
+  {
+    if (mem_check.condition)
+    {
+      gprs_used |= mem_check.condition->GetGPRsUsed();
+      fprs_used |= mem_check.condition->GetFPRsUsed();
+    }
+  }
+
+  const bool registers_changed =
+      gprs_used != m_gprs_used_in_conditions || fprs_used != m_fprs_used_in_conditions;
+
+  m_gprs_used_in_conditions = gprs_used;
+  m_fprs_used_in_conditions = fprs_used;
+
+  return registers_changed;
+}
+
 TMemCheck* MemChecks::GetMemCheck(u32 address, size_t size)
 {
   const auto iter = std::ranges::find_if(m_mem_checks, [address, size](const auto& mc) {

Source/Core/Core/PowerPC/BreakPoints.h

@@ -8,6 +8,7 @@
 #include <string>
 #include <vector>
 
+#include "Common/BitSet.h"
 #include "Common/CommonTypes.h"
 #include "Core/PowerPC/Expression.h"
 
@@ -127,8 +128,16 @@ public:
   void Clear();
   bool HasAny() const { return !m_mem_checks.empty(); }
 
+  BitSet32 GetGPRsUsedInConditions() { return m_gprs_used_in_conditions; }
+  BitSet32 GetFPRsUsedInConditions() { return m_fprs_used_in_conditions; }
+
 private:
+  // Returns whether any change was made
+  bool UpdateRegistersUsedInConditions();
+
   TMemChecks m_mem_checks;
   Core::System& m_system;
+  BitSet32 m_gprs_used_in_conditions;
+  BitSet32 m_fprs_used_in_conditions;
   bool m_mem_breakpoints_set = false;
 };

Source/Core/Core/PowerPC/Expression.cpp

@@ -494,3 +494,22 @@ std::string Expression::GetText() const
 {
   return m_text;
 }
+
+void Expression::ComputeRegistersUsed()
+{
+  if (m_has_computed_registers_used)
+    return;
+
+  for (const VarBinding& bind : m_binds)
+  {
+    switch (bind.type)
+    {
+    case VarBindingType::GPR:
+      m_gprs_used[bind.index] = true;
+      break;
+    case VarBindingType::FPR:
+      m_fprs_used[bind.index] = true;
+      break;
+    }
+  }
+}

Source/Core/Core/PowerPC/Expression.h

@@ -9,6 +9,8 @@
 #include <string_view>
 #include <vector>
 
+#include "Common/BitSet.h"
+
 struct expr;
 struct expr_var_list;
 
@@ -41,6 +43,18 @@ public:
 
   std::string GetText() const;
 
+  BitSet32 GetGPRsUsed()
+  {
+    ComputeRegistersUsed();
+    return m_gprs_used;
+  }
+
+  BitSet32 GetFPRsUsed()
+  {
+    ComputeRegistersUsed();
+    return m_fprs_used;
+  }
+
 private:
   enum class SynchronizeDirection
   {
@@ -69,10 +83,16 @@ private:
   void SynchronizeBindings(Core::System& system, SynchronizeDirection dir) const;
   void Reporting(const double result) const;
 
+  void ComputeRegistersUsed();
+
   std::string m_text;
   ExprPointer m_expr;
   ExprVarListPointer m_vars;
   std::vector<VarBinding> m_binds;
+
+  BitSet32 m_gprs_used;
+  BitSet32 m_fprs_used;
+  bool m_has_computed_registers_used = false;
 };
 
 inline bool EvaluateCondition(Core::System& system, const std::optional<Expression>& condition)

Source/Core/Core/PowerPC/Jit64/Jit.cpp

@@ -1294,6 +1294,32 @@ void Jit64::IntializeSpeculativeConstants()
   }
 }
 
+void Jit64::FlushPCBeforeSlowAccess()
+{
+  // PC is used by memory watchpoints (if enabled), profiling where to insert gather pipe
+  // interrupt checks, and printing accurate PC locations in debug logs.
+  MOV(32, PPCSTATE(pc), Imm32(js.compilerPC));
+}
+
+void Jit64::FlushRegistersBeforeSlowAccess()
+{
+  // Register values can be used by memory watchpoint conditions.
+  MemChecks& mem_checks = m_system.GetPowerPC().GetMemChecks();
+  if (mem_checks.HasAny())
+  {
+    BitSet32 gprs = mem_checks.GetGPRsUsedInConditions();
+    BitSet32 fprs = mem_checks.GetFPRsUsedInConditions();
+    if (gprs || fprs)
+    {
+      RCForkGuard gpr_guard = gpr.Fork();
+      RCForkGuard fpr_guard = fpr.Fork();
+
+      gpr.Flush(gprs);
+      fpr.Flush(fprs);
+    }
+  }
+}
+
 bool Jit64::HandleFunctionHooking(u32 address)
 {
   const auto result = HLE::TryReplaceFunction(m_ppc_symbol_db, address, PowerPC::CoreMode::JIT);

Source/Core/Core/PowerPC/Jit64/Jit.h

@@ -81,6 +81,9 @@ public:
 
   void IntializeSpeculativeConstants();
 
+  void FlushPCBeforeSlowAccess();
+  void FlushRegistersBeforeSlowAccess();
+
   JitBlockCache* GetBlockCache() override { return &blocks; }
   void Trace();
 

Source/Core/Core/PowerPC/Jit64/Jit_LoadStore.cpp

@@ -110,6 +110,8 @@ void Jit64::lXXx(UGeckoInstruction inst)
     PanicAlertFmt("Invalid instruction");
   }
 
+  FlushRegistersBeforeSlowAccess();
+
   // PowerPC has no 8-bit sign extended load, but x86 does, so merge extsb with the load if we find
   // it.
   if (CanMergeNextInstructions(1) && accessSize == 8 && js.op[1].inst.OPCD == 31 &&
@@ -439,6 +441,8 @@ void Jit64::dcbz(UGeckoInstruction inst)
   int a = inst.RA;
   int b = inst.RB;
 
+  FlushRegistersBeforeSlowAccess();
+
   {
     RCOpArg Ra = a ? gpr.Use(a, RCMode::Read) : RCOpArg::Imm32(0);
     RCOpArg Rb = gpr.Use(b, RCMode::Read);
@@ -477,7 +481,7 @@ void Jit64::dcbz(UGeckoInstruction inst)
     SwitchToFarCode();
     SetJumpTarget(slow);
   }
-  MOV(32, PPCSTATE(pc), Imm32(js.compilerPC));
+  FlushPCBeforeSlowAccess();
   BitSet32 registersInUse = CallerSavedRegistersInUse();
   ABI_PushRegistersAndAdjustStack(registersInUse, 0);
   ABI_CallFunctionPR(PowerPC::ClearDCacheLineFromJit, &m_mmu, RSCRATCH);
@@ -524,6 +528,8 @@ void Jit64::stX(UGeckoInstruction inst)
     return;
   }
 
+  FlushRegistersBeforeSlowAccess();
+
   // If we already know the address of the write
   if (!a || gpr.IsImm(a))
   {
@@ -602,6 +608,8 @@ void Jit64::stXx(UGeckoInstruction inst)
     break;
   }
 
+  FlushRegistersBeforeSlowAccess();
+
   const bool does_clobber = WriteClobbersRegValue(accessSize, /* swap */ !byte_reverse);
 
   RCOpArg Ra = update ? gpr.Bind(a, RCMode::ReadWrite) : gpr.Use(a, RCMode::Read);
@@ -634,6 +642,8 @@ void Jit64::lmw(UGeckoInstruction inst)
 
   int a = inst.RA, d = inst.RD;
 
+  FlushRegistersBeforeSlowAccess();
+
   // TODO: This doesn't handle rollback on DSI correctly
   {
     RCOpArg Ra = a ? gpr.Use(a, RCMode::Read) : RCOpArg::Imm32(0);
@@ -657,6 +667,8 @@ void Jit64::stmw(UGeckoInstruction inst)
 
   int a = inst.RA, d = inst.RD;
 
+  FlushRegistersBeforeSlowAccess();
+
   // TODO: This doesn't handle rollback on DSI correctly
   for (int i = d; i < 32; i++)
   {

Source/Core/Core/PowerPC/Jit64/Jit_LoadStoreFloating.cpp

@@ -30,6 +30,8 @@ void Jit64::lfXXX(UGeckoInstruction inst)
 
   FALLBACK_IF(!indexed && !a);
 
+  FlushRegistersBeforeSlowAccess();
+
   s32 offset = 0;
   RCOpArg addr = gpr.Bind(a, update ? RCMode::ReadWrite : RCMode::Read);
   RegCache::Realize(addr);
@@ -103,6 +105,8 @@ void Jit64::stfXXX(UGeckoInstruction inst)
 
   FALLBACK_IF(update && jo.memcheck && indexed && a == b);
 
+  FlushRegistersBeforeSlowAccess();
+
   if (single)
   {
     if (js.fpr_is_store_safe[s] && js.op->fprIsSingle[s])
@@ -196,6 +200,8 @@ void Jit64::stfiwx(UGeckoInstruction inst)
   int a = inst.RA;
   int b = inst.RB;
 
+  FlushRegistersBeforeSlowAccess();
+
   RCOpArg Ra = a ? gpr.Use(a, RCMode::Read) : RCOpArg::Imm32(0);
   RCOpArg Rb = gpr.Use(b, RCMode::Read);
   RCOpArg Rs = fpr.Use(s, RCMode::Read);

Source/Core/Core/PowerPC/Jit64/Jit_LoadStorePaired.cpp

@@ -35,6 +35,8 @@ void Jit64::psq_stXX(UGeckoInstruction inst)
   int w = indexed ? inst.Wx : inst.W;
   FALLBACK_IF(!a);
 
+  FlushRegistersBeforeSlowAccess();
+
   RCX64Reg scratch_guard = gpr.Scratch(RSCRATCH_EXTRA);
   RCOpArg Ra = update ? gpr.Bind(a, RCMode::ReadWrite) : gpr.Use(a, RCMode::Read);
   RCOpArg Rb = indexed ? gpr.Use(b, RCMode::Read) : RCOpArg::Imm32((u32)offset);
@@ -69,8 +71,8 @@ void Jit64::psq_stXX(UGeckoInstruction inst)
     }
     else
     {
-      // Stash PC in case asm routine needs to call into C++
+      FlushPCBeforeSlowAccess();
-      MOV(32, PPCSTATE(pc), Imm32(js.compilerPC));
+
       // We know what GQR is here, so we can load RSCRATCH2 and call into the store method directly
       // with just the scale bits.
       MOV(32, R(RSCRATCH2), Imm32(gqrValue & 0x3F00));
@@ -83,8 +85,8 @@ void Jit64::psq_stXX(UGeckoInstruction inst)
   }
   else
   {
-    // Stash PC in case asm routine needs to call into C++
+    FlushPCBeforeSlowAccess();
-    MOV(32, PPCSTATE(pc), Imm32(js.compilerPC));
+
     // Some games (e.g. Dirt 2) incorrectly set the unused bits which breaks the lookup table code.
     // Hence, we need to mask out the unused bits. The layout of the GQR register is
     // UU[SCALE]UUUUU[TYPE] where SCALE is 6 bits and TYPE is 3 bits, so we have to AND with
@@ -124,6 +126,8 @@ void Jit64::psq_lXX(UGeckoInstruction inst)
   int w = indexed ? inst.Wx : inst.W;
   FALLBACK_IF(!a);
 
+  FlushRegistersBeforeSlowAccess();
+
   RCX64Reg scratch_guard = gpr.Scratch(RSCRATCH_EXTRA);
   RCX64Reg Ra = gpr.Bind(a, update ? RCMode::ReadWrite : RCMode::Read);
   RCOpArg Rb = indexed ? gpr.Use(b, RCMode::Read) : RCOpArg::Imm32((u32)offset);
@@ -144,8 +148,8 @@ void Jit64::psq_lXX(UGeckoInstruction inst)
   }
   else
   {
-    // Stash PC in case asm routine needs to call into C++
+    FlushPCBeforeSlowAccess();
-    MOV(32, PPCSTATE(pc), Imm32(js.compilerPC));
+
     // Get the high part of the GQR register
     OpArg gqr = PPCSTATE_SPR(SPR_GQR0 + i);
     gqr.AddMemOffset(2);

Source/Core/Core/PowerPC/Jit64Common/EmuCodeBlock.cpp

@@ -386,15 +386,10 @@ void EmuCodeBlock::SafeLoadToReg(X64Reg reg_value, const Gen::OpArg& opAddress,
     SetJumpTarget(slow);
   }
 
-  // PC is used by memory watchpoints (if enabled), profiling where to insert gather pipe
+  // In the case of Jit64AsmCommon routines, the state we want to store here
-  // interrupt checks, and printing accurate PC locations in debug logs.
+  // isn't known at JIT time, so the caller has to store it themselves.
-  //
-  // In the case of Jit64AsmCommon routines, we don't know the PC here,
-  // so the caller has to store the PC themselves.
   if (!(flags & SAFE_LOADSTORE_NO_UPDATE_PC))
-  {
+    m_jit.FlushPCBeforeSlowAccess();
-    MOV(32, PPCSTATE(pc), Imm32(js.compilerPC));
-  }
 
   size_t rsp_alignment = (flags & SAFE_LOADSTORE_NO_PROLOG) ? 8 : 0;
   ABI_PushRegistersAndAdjustStack(registersInUse, rsp_alignment);
@@ -457,8 +452,7 @@ void EmuCodeBlock::SafeLoadToRegImmediate(X64Reg reg_value, u32 address, int acc
     return;
   }
 
-  // Helps external systems know which instruction triggered the read.
+  m_jit.FlushPCBeforeSlowAccess();
-  MOV(32, PPCSTATE(pc), Imm32(m_jit.js.compilerPC));
 
   // Fall back to general-case code.
   ABI_PushRegistersAndAdjustStack(registersInUse, 0);
@@ -560,15 +554,10 @@ void EmuCodeBlock::SafeWriteRegToReg(OpArg reg_value, X64Reg reg_addr, int acces
     SetJumpTarget(slow);
   }
 
-  // PC is used by memory watchpoints (if enabled), profiling where to insert gather pipe
+  // In the case of Jit64AsmCommon routines, the state we want to store here
-  // interrupt checks, and printing accurate PC locations in debug logs.
+  // isn't known at JIT time, so the caller has to store it themselves.
-  //
-  // In the case of Jit64AsmCommon routines, we don't know the PC here,
-  // so the caller has to store the PC themselves.
   if (!(flags & SAFE_LOADSTORE_NO_UPDATE_PC))
-  {
+    m_jit.FlushPCBeforeSlowAccess();
-    MOV(32, PPCSTATE(pc), Imm32(js.compilerPC));
-  }
 
   size_t rsp_alignment = (flags & SAFE_LOADSTORE_NO_PROLOG) ? 8 : 0;
   ABI_PushRegistersAndAdjustStack(registersInUse, rsp_alignment);
@@ -663,8 +652,7 @@ bool EmuCodeBlock::WriteToConstAddress(int accessSize, OpArg arg, u32 address,
   }
   else
   {
-    // Helps external systems know which instruction triggered the write
+    m_jit.FlushPCBeforeSlowAccess();
-    MOV(32, PPCSTATE(pc), Imm32(m_jit.js.compilerPC));
 
     ABI_PushRegistersAndAdjustStack(registersInUse, 0);
     switch (accessSize)

Source/Core/Core/PowerPC/JitArm64/Jit.h

@@ -273,12 +273,18 @@ protected:
   // !emitting_routine && mode != AlwaysSlowAccess && !jo.fastmem:                X30
   // !emitting_routine && mode == Auto && jo.fastmem:                             X30
   //
-  // Furthermore, any callee-saved register which isn't marked in gprs_to_push/fprs_to_push
+  // Furthermore:
-  // may be clobbered if mode != AlwaysFastAccess.
+  // - Any callee-saved register which isn't marked in gprs_to_push/fprs_to_push may be
+  //   clobbered if mode != AlwaysFastAccess.
+  // - If !emitting_routine && mode != AlwaysFastAccess && jo.memcheck, X30 must not
+  //   contain a guest register.
   void EmitBackpatchRoutine(u32 flags, MemAccessMode mode, Arm64Gen::ARM64Reg RS,
                             Arm64Gen::ARM64Reg addr, BitSet32 gprs_to_push = BitSet32(0),
                             BitSet32 fprs_to_push = BitSet32(0), bool emitting_routine = false);
 
+  // temp_gpr must be a valid register, but temp_fpr can be INVALID_REG.
+  void FlushPPCStateBeforeSlowAccess(Arm64Gen::ARM64Reg temp_gpr, Arm64Gen::ARM64Reg temp_fpr);
+
   // Loadstore routines
   void SafeLoadToReg(u32 dest, s32 addr, s32 offsetReg, u32 flags, s32 offset, bool update);
   void SafeStoreFromReg(s32 dest, u32 value, s32 regOffset, u32 flags, s32 offset, bool update);

Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp

@@ -21,6 +21,7 @@
 #include "Core/PowerPC/JitArmCommon/BackPatch.h"
 #include "Core/PowerPC/MMU.h"
 #include "Core/PowerPC/PowerPC.h"
+#include "Core/System.h"
 
 using namespace Arm64Gen;
 
@@ -171,6 +172,7 @@ void JitArm64::EmitBackpatchRoutine(u32 flags, MemAccessMode mode, ARM64Reg RS,
 
     const ARM64Reg temp_gpr = ARM64Reg::W1;
     const int temp_gpr_index = DecodeReg(temp_gpr);
+    const ARM64Reg temp_fpr = fprs_to_push[0] ? ARM64Reg::INVALID_REG : ARM64Reg::Q0;
 
     BitSet32 gprs_to_push_early = {};
     if (memcheck)
@@ -189,16 +191,10 @@ void JitArm64::EmitBackpatchRoutine(u32 flags, MemAccessMode mode, ARM64Reg RS,
     ABI_PushRegisters(gprs_to_push & ~gprs_to_push_early);
     m_float_emit.ABI_PushRegisters(fprs_to_push, ARM64Reg::X30);
 
-    // PC is used by memory watchpoints (if enabled), profiling where to insert gather pipe
+    // In the case of JitAsm routines, the state we want to store here
-    // interrupt checks, and printing accurate PC locations in debug logs.
+    // isn't known at JIT time, so the caller has to store it themselves.
-    //
-    // In the case of JitAsm routines, we don't know the PC here,
-    // so the caller has to store the PC themselves.
     if (!emitting_routine)
-    {
+      FlushPPCStateBeforeSlowAccess(ARM64Reg::W30, temp_fpr);
-      MOVI2R(ARM64Reg::W30, js.compilerPC);
-      STR(IndexType::Unsigned, ARM64Reg::W30, PPC_REG, PPCSTATE_OFF(pc));
-    }
 
     if (flags & BackPatchInfo::FLAG_STORE)
     {
@@ -265,7 +261,6 @@ void JitArm64::EmitBackpatchRoutine(u32 flags, MemAccessMode mode, ARM64Reg RS,
 
     if (memcheck)
     {
-      const ARM64Reg temp_fpr = fprs_to_push[0] ? ARM64Reg::INVALID_REG : ARM64Reg::Q0;
       const u64 early_push_count = (gprs_to_push & gprs_to_push_early).Count();
       const u64 early_push_size = Common::AlignUp(early_push_count, 2) * 8;
 
@@ -316,6 +311,22 @@ void JitArm64::EmitBackpatchRoutine(u32 flags, MemAccessMode mode, ARM64Reg RS,
   }
 }
 
+void JitArm64::FlushPPCStateBeforeSlowAccess(ARM64Reg temp_gpr, ARM64Reg temp_fpr)
+{
+  // PC is used by memory watchpoints (if enabled), profiling where to insert gather pipe
+  // interrupt checks, and printing accurate PC locations in debug logs.
+  MOVI2R(temp_gpr, js.compilerPC);
+  STR(IndexType::Unsigned, temp_gpr, PPC_REG, PPCSTATE_OFF(pc));
+
+  // Register values can be used by memory watchpoint conditions.
+  MemChecks& mem_checks = m_system.GetPowerPC().GetMemChecks();
+  if (mem_checks.HasAny())
+  {
+    gpr.StoreRegisters(mem_checks.GetGPRsUsedInConditions(), temp_gpr, FlushMode::MaintainState);
+    fpr.StoreRegisters(mem_checks.GetFPRsUsedInConditions(), temp_fpr, FlushMode::MaintainState);
+  }
+}
+
 bool JitArm64::HandleFastmemFault(SContext* ctx)
 {
   const u8* pc = reinterpret_cast<const u8*>(ctx->CTX_PC);

Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStorePaired.cpp

@@ -102,9 +102,7 @@ void JitArm64::psq_lXX(UGeckoInstruction inst)
   {
     LDR(IndexType::Unsigned, scale_reg, PPC_REG, PPCSTATE_OFF_SPR(SPR_GQR0 + i));
 
-    // Stash PC in case asm routine needs to call into C++
+    FlushPPCStateBeforeSlowAccess(ARM64Reg::W30, ARM64Reg::Q1);
-    MOVI2R(ARM64Reg::W30, js.compilerPC);
-    STR(IndexType::Unsigned, ARM64Reg::W30, PPC_REG, PPCSTATE_OFF(pc));
 
     UBFM(type_reg, scale_reg, 16, 18);   // Type
     UBFM(scale_reg, scale_reg, 24, 29);  // Scale
@@ -260,9 +258,7 @@ void JitArm64::psq_stXX(UGeckoInstruction inst)
   {
     LDR(IndexType::Unsigned, scale_reg, PPC_REG, PPCSTATE_OFF_SPR(SPR_GQR0 + i));
 
-    // Stash PC in case asm routine needs to call into C++
+    FlushPPCStateBeforeSlowAccess(ARM64Reg::W30, ARM64Reg::Q1);
-    MOVI2R(ARM64Reg::W30, js.compilerPC);
-    STR(IndexType::Unsigned, ARM64Reg::W30, PPC_REG, PPCSTATE_OFF(pc));
 
     UBFM(type_reg, scale_reg, 0, 2);    // Type
     UBFM(scale_reg, scale_reg, 8, 13);  // Scale

Source/Core/Core/PowerPC/JitArm64/JitArm64_RegCache.h

@@ -388,14 +388,16 @@ public:
 
   BitSet32 GetDirtyGPRs() const;
 
-  void StoreRegisters(BitSet32 regs, Arm64Gen::ARM64Reg tmp_reg = Arm64Gen::ARM64Reg::INVALID_REG)
+  void StoreRegisters(BitSet32 regs, Arm64Gen::ARM64Reg tmp_reg = Arm64Gen::ARM64Reg::INVALID_REG,
+                      FlushMode flush_mode = FlushMode::All)
   {
-    FlushRegisters(regs, FlushMode::All, tmp_reg, IgnoreDiscardedRegisters::No);
+    FlushRegisters(regs, flush_mode, tmp_reg, IgnoreDiscardedRegisters::No);
   }
 
-  void StoreCRRegisters(BitSet8 regs, Arm64Gen::ARM64Reg tmp_reg = Arm64Gen::ARM64Reg::INVALID_REG)
+  void StoreCRRegisters(BitSet8 regs, Arm64Gen::ARM64Reg tmp_reg = Arm64Gen::ARM64Reg::INVALID_REG,
+                        FlushMode flush_mode = FlushMode::All)
   {
-    FlushCRRegisters(regs, FlushMode::All, tmp_reg, IgnoreDiscardedRegisters::No);
+    FlushCRRegisters(regs, flush_mode, tmp_reg, IgnoreDiscardedRegisters::No);
   }
 
   void DiscardCRRegisters(BitSet8 regs);
@@ -459,9 +461,10 @@ public:
 
   void FixSinglePrecision(size_t preg);
 
-  void StoreRegisters(BitSet32 regs, Arm64Gen::ARM64Reg tmp_reg = Arm64Gen::ARM64Reg::INVALID_REG)
+  void StoreRegisters(BitSet32 regs, Arm64Gen::ARM64Reg tmp_reg = Arm64Gen::ARM64Reg::INVALID_REG,
+                      FlushMode flush_mode = FlushMode::All)
   {
-    FlushRegisters(regs, FlushMode::All, tmp_reg);
+    FlushRegisters(regs, flush_mode, tmp_reg);
   }
 
 protected: