diff --git a/src/core/cpu_newrec_compiler_aarch32.cpp b/src/core/cpu_newrec_compiler_aarch32.cpp
index d3f97b97e..10a009f43 100644
--- a/src/core/cpu_newrec_compiler_aarch32.cpp
+++ b/src/core/cpu_newrec_compiler_aarch32.cpp
@@ -1576,6 +1576,8 @@ void CPU::NewRec::AArch32Compiler::Compile_lwx(CompileFlags cf, MemoryAccessSize
                                                const std::optional<VirtualMemoryAddress>& address)
 {
   DebugAssert(size == MemoryAccessSize::Word && !sign);
+
+  const Register addr = Register(AllocateTempHostReg(HR_CALLEE_SAVED));
   FlushForLoadStore(address, false, use_fastmem);
 
   // TODO: if address is constant, this can be simplified..
@@ -1585,7 +1587,6 @@ void CPU::NewRec::AArch32Compiler::Compile_lwx(CompileFlags cf, MemoryAccessSize
     UpdateLoadDelay();
 
   // We'd need to be careful here if we weren't overwriting it..
-  const Register addr = Register(AllocateHostReg(HR_CALLEE_SAVED, HR_TYPE_TEMP));
   ComputeLoadStoreAddressArg(cf, address, addr);
   armAsm->and_(RARG1, addr, armCheckLogicalConstant(~0x3u));
   GenerateLoad(RARG1, MemoryAccessSize::Word, false, use_fastmem, []() { return RRET; });
@@ -1768,11 +1769,12 @@ void CPU::NewRec::AArch32Compiler::Compile_swx(CompileFlags cf, MemoryAccessSize
                                                const std::optional<VirtualMemoryAddress>& address)
 {
   DebugAssert(size == MemoryAccessSize::Word && !sign);
+
+  const Register addr = Register(AllocateTempHostReg(HR_CALLEE_SAVED));
   FlushForLoadStore(address, true, use_fastmem);
 
   // TODO: if address is constant, this can be simplified..
   // We'd need to be careful here if we weren't overwriting it..
-  const Register addr = Register(AllocateHostReg(HR_CALLEE_SAVED, HR_TYPE_TEMP));
   ComputeLoadStoreAddressArg(cf, address, addr);
   armAsm->and_(RARG1, addr, armCheckLogicalConstant(~0x3u));
   GenerateLoad(RARG1, MemoryAccessSize::Word, false, use_fastmem, []() { return RRET; });
diff --git a/src/core/cpu_newrec_compiler_aarch64.cpp b/src/core/cpu_newrec_compiler_aarch64.cpp
index 2840fea29..a30d9133f 100644
--- a/src/core/cpu_newrec_compiler_aarch64.cpp
+++ b/src/core/cpu_newrec_compiler_aarch64.cpp
@@ -1555,6 +1555,8 @@ void CPU::NewRec::AArch64Compiler::Compile_lwx(CompileFlags cf, MemoryAccessSize
                                                const std::optional<VirtualMemoryAddress>& address)
 {
   DebugAssert(size == MemoryAccessSize::Word && !sign);
+
+  const WRegister addr = WRegister(AllocateTempHostReg(HR_CALLEE_SAVED));
   FlushForLoadStore(address, false, use_fastmem);
 
   // TODO: if address is constant, this can be simplified..
@@ -1564,7 +1566,6 @@ void CPU::NewRec::AArch64Compiler::Compile_lwx(CompileFlags cf, MemoryAccessSize
     UpdateLoadDelay();
 
   // We'd need to be careful here if we weren't overwriting it..
-  const WRegister addr = WRegister(AllocateHostReg(HR_CALLEE_SAVED, HR_TYPE_TEMP));
   ComputeLoadStoreAddressArg(cf, address, addr);
   armAsm->and_(RWARG1, addr, armCheckLogicalConstant(~0x3u));
   GenerateLoad(RWARG1, MemoryAccessSize::Word, false, use_fastmem, []() { return RWRET; });
@@ -1747,11 +1748,12 @@ void CPU::NewRec::AArch64Compiler::Compile_swx(CompileFlags cf, MemoryAccessSize
                                                const std::optional<VirtualMemoryAddress>& address)
 {
   DebugAssert(size == MemoryAccessSize::Word && !sign);
+
+  const WRegister addr = WRegister(AllocateTempHostReg(HR_CALLEE_SAVED));
   FlushForLoadStore(address, true, use_fastmem);
 
   // TODO: if address is constant, this can be simplified..
   // We'd need to be careful here if we weren't overwriting it..
-  const WRegister addr = WRegister(AllocateHostReg(HR_CALLEE_SAVED, HR_TYPE_TEMP));
   ComputeLoadStoreAddressArg(cf, address, addr);
   armAsm->and_(RWARG1, addr, armCheckLogicalConstant(~0x3u));
   GenerateLoad(RWARG1, MemoryAccessSize::Word, false, use_fastmem, []() { return RWRET; });
diff --git a/src/core/cpu_newrec_compiler_riscv64.cpp b/src/core/cpu_newrec_compiler_riscv64.cpp
index 5c2f646c1..d5d52295b 100644
--- a/src/core/cpu_newrec_compiler_riscv64.cpp
+++ b/src/core/cpu_newrec_compiler_riscv64.cpp
@@ -1847,6 +1847,8 @@ void CPU::NewRec::RISCV64Compiler::Compile_lwx(CompileFlags cf, MemoryAccessSize
                                                const std::optional<VirtualMemoryAddress>& address)
 {
   DebugAssert(size == MemoryAccessSize::Word && !sign);
+
+  const GPR addr = GPR(AllocateTempHostReg(HR_CALLEE_SAVED));
   FlushForLoadStore(address, false, use_fastmem);
 
   // TODO: if address is constant, this can be simplified..
@@ -1856,7 +1858,6 @@ void CPU::NewRec::RISCV64Compiler::Compile_lwx(CompileFlags cf, MemoryAccessSize
     UpdateLoadDelay();
 
   // We'd need to be careful here if we weren't overwriting it..
-  const GPR addr = GPR(AllocateHostReg(HR_CALLEE_SAVED, HR_TYPE_TEMP));
   ComputeLoadStoreAddressArg(cf, address, addr);
   rvAsm->ANDI(RARG1, addr, ~0x3u);
   GenerateLoad(RARG1, MemoryAccessSize::Word, false, use_fastmem, []() { return RRET; });
@@ -2012,11 +2013,12 @@ void CPU::NewRec::RISCV64Compiler::Compile_swx(CompileFlags cf, MemoryAccessSize
                                                const std::optional<VirtualMemoryAddress>& address)
 {
   DebugAssert(size == MemoryAccessSize::Word && !sign);
+
+  const GPR addr = GPR(AllocateTempHostReg(HR_CALLEE_SAVED));
   FlushForLoadStore(address, true, use_fastmem);
 
   // TODO: if address is constant, this can be simplified..
   // We'd need to be careful here if we weren't overwriting it..
-  const GPR addr = GPR(AllocateHostReg(HR_CALLEE_SAVED, HR_TYPE_TEMP));
   ComputeLoadStoreAddressArg(cf, address, addr);
   rvAsm->ANDI(RARG1, addr, ~0x3u);
   GenerateLoad(RARG1, MemoryAccessSize::Word, false, use_fastmem, []() { return RRET; });
diff --git a/src/core/cpu_newrec_compiler_x64.cpp b/src/core/cpu_newrec_compiler_x64.cpp
index 5d5629a22..0b96f8eef 100644
--- a/src/core/cpu_newrec_compiler_x64.cpp
+++ b/src/core/cpu_newrec_compiler_x64.cpp
@@ -1515,6 +1515,8 @@ void CPU::NewRec::X64Compiler::Compile_lwx(CompileFlags cf, MemoryAccessSize siz
                                            const std::optional<VirtualMemoryAddress>& address)
 {
   DebugAssert(size == MemoryAccessSize::Word && !sign);
+
+  const Reg32 addr = Reg32(AllocateTempHostReg(HR_CALLEE_SAVED));
   FlushForLoadStore(address, false, use_fastmem);
 
   // TODO: if address is constant, this can be simplified..
@@ -1524,7 +1526,6 @@ void CPU::NewRec::X64Compiler::Compile_lwx(CompileFlags cf, MemoryAccessSize siz
     UpdateLoadDelay();
 
   // We'd need to be careful here if we weren't overwriting it..
-  const Reg32 addr = Reg32(AllocateHostReg(HR_CALLEE_SAVED, HR_TYPE_TEMP));
   ComputeLoadStoreAddressArg(cf, address, addr);
   cg->mov(RWARG1, addr);
   cg->and_(RWARG1, ~0x3u);
@@ -1711,11 +1712,12 @@ void CPU::NewRec::X64Compiler::Compile_swx(CompileFlags cf, MemoryAccessSize siz
                                            const std::optional<VirtualMemoryAddress>& address)
 {
   DebugAssert(size == MemoryAccessSize::Word && !sign);
+
+  const Reg32 addr = Reg32(AllocateTempHostReg(HR_CALLEE_SAVED));
   FlushForLoadStore(address, true, use_fastmem);
 
   // TODO: if address is constant, this can be simplified..
   // We'd need to be careful here if we weren't overwriting it..
-  const Reg32 addr = Reg32(AllocateHostReg(HR_CALLEE_SAVED, HR_TYPE_TEMP));
   ComputeLoadStoreAddressArg(cf, address, addr);
   cg->mov(RWARG1, addr);
   cg->and_(RWARG1, ~0x3u);