CPU/NewRec: Add AArch32 backend

CMakeLists.txt

@@ -23,7 +23,6 @@ endif()
 # Renderer options.
 option(ENABLE_OPENGL "Build with OpenGL renderer" ON)
 option(ENABLE_VULKAN "Build with Vulkan renderer" ON)
-option(ENABLE_NEWREC "Build with experimental new dynarec (needed for RISC-V)" ON)
 
 # Global options.
 if(NOT ANDROID)

dep/vixl/include/vixl/code-buffer-vixl.h

@@ -43,7 +43,7 @@ class CodeBuffer {
   ~CodeBuffer();
 
   void Reset();
-  void Reset(byte* buffer, size_t capacity);
+  void Reset(byte* buffer, size_t capacity, bool managed = false);
 
 #ifdef VIXL_CODE_BUFFER_MMAP
   void SetExecutable();

dep/vixl/src/code-buffer-vixl.cc

@@ -156,10 +156,11 @@ void CodeBuffer::Reset() {
   SetClean();
 }
 
-void CodeBuffer::Reset(byte* buffer, size_t capacity) {
+void CodeBuffer::Reset(byte* buffer, size_t capacity, bool managed) {
   buffer_ = buffer;
   cursor_ = buffer;
   capacity_ = capacity;
+  managed_ = managed;
 }
 
 void CodeBuffer::Grow(size_t new_capacity) {

dep/xbyak/CMakeLists.txt

@@ -1,2 +1,3 @@
 add_library(xbyak INTERFACE)
 target_include_directories(xbyak INTERFACE "${CMAKE_CURRENT_SOURCE_DIR}/xbyak")
+target_compile_definitions(xbyak INTERFACE "XBYAK_NO_EXCEPTION=1")

src/core/CMakeLists.txt

@@ -133,54 +133,40 @@ target_link_libraries(core PUBLIC Threads::Threads common util zlib)
 target_link_libraries(core PRIVATE stb xxhash imgui rapidjson rcheevos)
 
 if(${CPU_ARCH} STREQUAL "x64")
-  target_compile_definitions(core PUBLIC "XBYAK_NO_EXCEPTION=1" "ENABLE_RECOMPILER=1" "ENABLE_MMAP_FASTMEM=1")
-  target_sources(core PRIVATE ${RECOMPILER_SRCS}
+  target_compile_definitions(core PUBLIC "ENABLE_RECOMPILER=1" "ENABLE_NEWREC=1" "ENABLE_MMAP_FASTMEM=1")
+  target_sources(core PRIVATE ${RECOMPILER_SRCS} ${NEWREC_SOURCES}
     cpu_recompiler_code_generator_x64.cpp
+    cpu_newrec_compiler_x64.cpp
+    cpu_newrec_compiler_x64.h
   )
   target_link_libraries(core PRIVATE xbyak zydis)
   message("Building x64 recompiler")
-
-  if(ENABLE_NEWREC)
-    target_compile_definitions(core PUBLIC "ENABLE_NEWREC=1")
-    target_sources(core PRIVATE ${NEWREC_SOURCES}
-      cpu_newrec_compiler_x64.cpp
-      cpu_newrec_compiler_x64.h
-    )
-    message("Building x64 newrec")
-  endif()
 elseif(${CPU_ARCH} STREQUAL "aarch32")
-  target_compile_definitions(core PUBLIC "ENABLE_RECOMPILER=1")
-  target_sources(core PRIVATE ${RECOMPILER_SRCS}
+  target_compile_definitions(core PUBLIC "ENABLE_RECOMPILER=1" "ENABLE_NEWREC=1")
+  target_sources(core PRIVATE ${RECOMPILER_SRCS} ${NEWREC_SOURCES}
     cpu_recompiler_code_generator_aarch32.cpp
+    cpu_newrec_compiler_aarch32.cpp
+    cpu_newrec_compiler_aarch32.h
   )
   target_link_libraries(core PUBLIC vixl)
   message("Building AArch32 recompiler")
 elseif(${CPU_ARCH} STREQUAL "aarch64")
-  target_compile_definitions(core PUBLIC "ENABLE_RECOMPILER=1" "ENABLE_MMAP_FASTMEM=1")
-  target_sources(core PRIVATE ${RECOMPILER_SRCS}
+  target_compile_definitions(core PUBLIC "ENABLE_RECOMPILER=1" "ENABLE_NEWREC=1" "ENABLE_MMAP_FASTMEM=1")
+  target_sources(core PRIVATE ${RECOMPILER_SRCS} ${NEWREC_SOURCES}
     cpu_recompiler_code_generator_aarch64.cpp
+    cpu_newrec_compiler_aarch64.cpp
+    cpu_newrec_compiler_aarch64.h
   )
   target_link_libraries(core PUBLIC vixl)
   message("Building AArch64 recompiler")
-  if(ENABLE_NEWREC)
-    target_compile_definitions(core PUBLIC "ENABLE_NEWREC=1")
-    target_sources(core PRIVATE ${NEWREC_SOURCES}
-      cpu_newrec_compiler_aarch64.cpp
-      cpu_newrec_compiler_aarch64.h
-    )
-    message("Building AArch64 newrec")
-  endif()
 elseif(${CPU_ARCH} STREQUAL "riscv64")
-  target_compile_definitions(core PUBLIC "ENABLE_MMAP_FASTMEM=1")
-  if(ENABLE_NEWREC)
-    target_compile_definitions(core PUBLIC "ENABLE_NEWREC=1")
-    target_sources(core PRIVATE ${NEWREC_SOURCES}
-      cpu_newrec_compiler_riscv64.cpp
-      cpu_newrec_compiler_riscv64.h
-    )
-    target_link_libraries(core PUBLIC biscuit::biscuit riscv-disas)
-    message("Building RISC-V 64-bit newrec")
-  endif()
+  target_compile_definitions(core PUBLIC "ENABLE_NEWREC=1" "ENABLE_MMAP_FASTMEM=1")
+  target_sources(core PRIVATE ${NEWREC_SOURCES}
+    cpu_newrec_compiler_riscv64.cpp
+    cpu_newrec_compiler_riscv64.h
+  )
+  target_link_libraries(core PUBLIC biscuit::biscuit riscv-disas)
+  message("Building RISC-V 64-bit recompiler")
 else()
   message("Not building recompiler")
 endif()

src/core/core.vcxproj

@@ -14,9 +14,15 @@
     <ClCompile Include="cpu_disasm.cpp" />
     <ClCompile Include="cpu_code_cache.cpp" />
     <ClCompile Include="cpu_newrec_compiler.cpp" />
+    <ClCompile Include="cpu_newrec_compiler_aarch32.cpp">
+      <ExcludedFromBuild Condition="'$(Platform)'!='ARM'">true</ExcludedFromBuild>
+    </ClCompile>
     <ClCompile Include="cpu_newrec_compiler_aarch64.cpp">
       <ExcludedFromBuild Condition="'$(Platform)'!='ARM64'">true</ExcludedFromBuild>
     </ClCompile>
+    <ClCompile Include="cpu_newrec_compiler_riscv64.cpp">
+      <ExcludedFromBuild>true</ExcludedFromBuild>
+    </ClCompile>
     <ClCompile Include="cpu_newrec_compiler_x64.cpp">
       <ExcludedFromBuild Condition="'$(Platform)'!='x64'">true</ExcludedFromBuild>
     </ClCompile>
@@ -98,9 +104,15 @@
     <ClInclude Include="cpu_disasm.h" />
     <ClInclude Include="cpu_code_cache.h" />
     <ClInclude Include="cpu_newrec_compiler.h" />
+    <ClInclude Include="cpu_newrec_compiler_aarch32.h">
+      <ExcludedFromBuild Condition="'$(Platform)'!='ARM'">true</ExcludedFromBuild>
+    </ClInclude>
     <ClInclude Include="cpu_newrec_compiler_aarch64.h">
       <ExcludedFromBuild Condition="'$(Platform)'!='ARM64'">true</ExcludedFromBuild>
     </ClInclude>
+    <ClInclude Include="cpu_newrec_compiler_riscv64.h">
+      <ExcludedFromBuild>true</ExcludedFromBuild>
+    </ClInclude>
     <ClInclude Include="cpu_newrec_compiler_x64.h">
       <ExcludedFromBuild Condition="'$(Platform)'!='x64'">true</ExcludedFromBuild>
     </ClInclude>

src/core/core.vcxproj.filters

@@ -63,6 +63,8 @@
     <ClCompile Include="cpu_newrec_compiler.cpp" />
     <ClCompile Include="cpu_newrec_compiler_x64.cpp" />
     <ClCompile Include="cpu_newrec_compiler_aarch64.cpp" />
+    <ClCompile Include="cpu_newrec_compiler_riscv64.cpp" />
+    <ClCompile Include="cpu_newrec_compiler_aarch32.cpp" />
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="types.h" />
@@ -131,5 +133,7 @@
     <ClInclude Include="cpu_newrec_compiler.h" />
     <ClInclude Include="cpu_newrec_compiler_x64.h" />
     <ClInclude Include="cpu_newrec_compiler_aarch64.h" />
+    <ClInclude Include="cpu_newrec_compiler_riscv64.h" />
+    <ClInclude Include="cpu_newrec_compiler_aarch32.h" />
   </ItemGroup>
 </Project>

src/core/cpu_newrec_compiler.h

@@ -21,6 +21,9 @@ static constexpr bool SWAP_BRANCH_DELAY_SLOTS = true;
 #if defined(CPU_ARCH_X64)
 static constexpr u32 NUM_HOST_REGS = 16;
 static constexpr bool HAS_MEMORY_OPERANDS = true;
+#elif defined(CPU_ARCH_ARM32)
+static constexpr u32 NUM_HOST_REGS = 16;
+static constexpr bool HAS_MEMORY_OPERANDS = false;
 #elif defined(CPU_ARCH_ARM64)
 static constexpr u32 NUM_HOST_REGS = 32;
 static constexpr bool HAS_MEMORY_OPERANDS = false;
@@ -153,6 +156,7 @@ protected:
     HR_TYPE_PC_WRITEBACK,
     HR_TYPE_LOAD_DELAY_VALUE,
     HR_TYPE_NEXT_LOAD_DELAY_VALUE,
+    HR_TYPE_MEMBASE,
   };
 
   struct HostRegAlloc

src/core/cpu_newrec_compiler_aarch32.h

@@ -0,0 +1,164 @@
+// SPDX-FileCopyrightText: 2023 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
+
+#pragma once
+#include "cpu_newrec_compiler.h"
+#include <memory>
+
+#include "vixl/aarch32/assembler-aarch32.h"
+#include "vixl/aarch32/operands-aarch32.h"
+
+namespace CPU::NewRec {
+
+class AArch32Compiler final : public Compiler
+{
+public:
+  AArch32Compiler();
+  ~AArch32Compiler() override;
+
+protected:
+  const char* GetHostRegName(u32 reg) const override;
+
+  const void* GetCurrentCodePointer() override;
+
+  void LoadHostRegWithConstant(u32 reg, u32 val) override;
+  void LoadHostRegFromCPUPointer(u32 reg, const void* ptr) override;
+  void StoreConstantToCPUPointer(u32 val, const void* ptr) override;
+  void StoreHostRegToCPUPointer(u32 reg, const void* ptr) override;
+  void CopyHostReg(u32 dst, u32 src) override;
+
+  void Reset(CodeCache::Block* block, u8* code_buffer, u32 code_buffer_space, u8* far_code_buffer,
+             u32 far_code_space) override;
+  void BeginBlock() override;
+  void GenerateBlockProtectCheck(const u8* ram_ptr, const u8* shadow_ptr, u32 size) override;
+  void GenerateICacheCheckAndUpdate() override;
+  void GenerateCall(const void* func, s32 arg1reg = -1, s32 arg2reg = -1, s32 arg3reg = -1) override;
+  void EndBlock(const std::optional<u32>& newpc, bool do_event_test) override;
+  void EndBlockWithException(Exception excode) override;
+  void EndAndLinkBlock(const std::optional<u32>& newpc, bool do_event_test);
+  const void* EndCompile(u32* code_size, u32* far_code_size) override;
+
+  void Flush(u32 flags) override;
+
+  void Compile_Fallback() override;
+
+  void CheckBranchTarget(const vixl::aarch32::Register& pcreg);
+  void Compile_jr(CompileFlags cf) override;
+  void Compile_jalr(CompileFlags cf) override;
+  void Compile_bxx(CompileFlags cf, BranchCondition cond) override;
+
+  void Compile_addi(CompileFlags cf, bool overflow);
+  void Compile_addi(CompileFlags cf) override;
+  void Compile_addiu(CompileFlags cf) override;
+  void Compile_slti(CompileFlags cf, bool sign);
+  void Compile_slti(CompileFlags cf) override;
+  void Compile_sltiu(CompileFlags cf) override;
+  void Compile_andi(CompileFlags cf) override;
+  void Compile_ori(CompileFlags cf) override;
+  void Compile_xori(CompileFlags cf) override;
+
+  void Compile_shift(CompileFlags cf,
+                     void (vixl::aarch32::Assembler::*op)(vixl::aarch32::Register, vixl::aarch32::Register,
+                                                          const vixl::aarch32::Operand&));
+  void Compile_sll(CompileFlags cf) override;
+  void Compile_srl(CompileFlags cf) override;
+  void Compile_sra(CompileFlags cf) override;
+  void Compile_variable_shift(CompileFlags cf,
+                              void (vixl::aarch32::Assembler::*op)(vixl::aarch32::Register, vixl::aarch32::Register,
+                                                                   const vixl::aarch32::Operand&));
+  void Compile_sllv(CompileFlags cf) override;
+  void Compile_srlv(CompileFlags cf) override;
+  void Compile_srav(CompileFlags cf) override;
+  void Compile_mult(CompileFlags cf, bool sign);
+  void Compile_mult(CompileFlags cf) override;
+  void Compile_multu(CompileFlags cf) override;
+  void Compile_div(CompileFlags cf) override;
+  void Compile_divu(CompileFlags cf) override;
+  void TestOverflow(const vixl::aarch32::Register& result);
+  void Compile_dst_op(CompileFlags cf,
+                      void (vixl::aarch32::Assembler::*op)(vixl::aarch32::Register, vixl::aarch32::Register,
+                                                           const vixl::aarch32::Operand&),
+                      bool commutative, bool logical, bool overflow);
+  void Compile_add(CompileFlags cf) override;
+  void Compile_addu(CompileFlags cf) override;
+  void Compile_sub(CompileFlags cf) override;
+  void Compile_subu(CompileFlags cf) override;
+  void Compile_and(CompileFlags cf) override;
+  void Compile_or(CompileFlags cf) override;
+  void Compile_xor(CompileFlags cf) override;
+  void Compile_nor(CompileFlags cf) override;
+  void Compile_slt(CompileFlags cf, bool sign);
+  void Compile_slt(CompileFlags cf) override;
+  void Compile_sltu(CompileFlags cf) override;
+
+  vixl::aarch32::Register
+  ComputeLoadStoreAddressArg(CompileFlags cf, const std::optional<VirtualMemoryAddress>& address,
+                             const std::optional<const vixl::aarch32::Register>& reg = std::nullopt);
+  template<typename RegAllocFn>
+  vixl::aarch32::Register GenerateLoad(const vixl::aarch32::Register& addr_reg, MemoryAccessSize size, bool sign,
+                                       const RegAllocFn& dst_reg_alloc);
+  void GenerateStore(const vixl::aarch32::Register& addr_reg, const vixl::aarch32::Register& value_reg,
+                     MemoryAccessSize size);
+  void Compile_lxx(CompileFlags cf, MemoryAccessSize size, bool sign,
+                   const std::optional<VirtualMemoryAddress>& address) override;
+  void Compile_lwx(CompileFlags cf, MemoryAccessSize size, bool sign,
+                   const std::optional<VirtualMemoryAddress>& address) override;
+  void Compile_lwc2(CompileFlags cf, MemoryAccessSize size, bool sign,
+                    const std::optional<VirtualMemoryAddress>& address) override;
+  void Compile_sxx(CompileFlags cf, MemoryAccessSize size, bool sign,
+                   const std::optional<VirtualMemoryAddress>& address) override;
+  void Compile_swx(CompileFlags cf, MemoryAccessSize size, bool sign,
+                   const std::optional<VirtualMemoryAddress>& address) override;
+  void Compile_swc2(CompileFlags cf, MemoryAccessSize size, bool sign,
+                    const std::optional<VirtualMemoryAddress>& address) override;
+
+  void TestInterrupts(const vixl::aarch32::Register& sr);
+  void Compile_mtc0(CompileFlags cf) override;
+  void Compile_rfe(CompileFlags cf) override;
+
+  void Compile_mfc2(CompileFlags cf) override;
+  void Compile_mtc2(CompileFlags cf) override;
+  void Compile_cop2(CompileFlags cf) override;
+
+  void GeneratePGXPCallWithMIPSRegs(const void* func, u32 arg1val, Reg arg2reg = Reg::count,
+                                    Reg arg3reg = Reg::count) override;
+
+private:
+  void EmitMov(const vixl::aarch32::Register& dst, u32 val);
+  void EmitCall(const void* ptr, bool force_inline = false);
+
+  vixl::aarch32::Operand armCheckAddSubConstant(s32 val);
+  vixl::aarch32::Operand armCheckAddSubConstant(u32 val);
+  vixl::aarch32::Operand armCheckCompareConstant(s32 val);
+  vixl::aarch32::Operand armCheckLogicalConstant(u32 val);
+
+  void SwitchToFarCode(bool emit_jump, vixl::aarch32::ConditionType cond = vixl::aarch32::ConditionType::al);
+  void SwitchToFarCodeIfBitSet(const vixl::aarch32::Register& reg, u32 bit);
+  void SwitchToFarCodeIfRegZeroOrNonZero(const vixl::aarch32::Register& reg, bool nonzero);
+  void SwitchToNearCode(bool emit_jump, vixl::aarch32::ConditionType cond = vixl::aarch32::ConditionType::al);
+
+  void AssertRegOrConstS(CompileFlags cf) const;
+  void AssertRegOrConstT(CompileFlags cf) const;
+  vixl::aarch32::MemOperand MipsPtr(Reg r) const;
+  vixl::aarch32::Register CFGetRegD(CompileFlags cf) const;
+  vixl::aarch32::Register CFGetRegS(CompileFlags cf) const;
+  vixl::aarch32::Register CFGetRegT(CompileFlags cf) const;
+  vixl::aarch32::Register CFGetRegLO(CompileFlags cf) const;
+  vixl::aarch32::Register CFGetRegHI(CompileFlags cf) const;
+  vixl::aarch32::Register GetMembaseReg();
+
+  void MoveSToReg(const vixl::aarch32::Register& dst, CompileFlags cf);
+  void MoveTToReg(const vixl::aarch32::Register& dst, CompileFlags cf);
+  void MoveMIPSRegToReg(const vixl::aarch32::Register& dst, Reg reg);
+
+  vixl::aarch32::Assembler m_emitter;
+  vixl::aarch32::Assembler m_far_emitter;
+  vixl::aarch32::Assembler* armAsm;
+
+#ifdef VIXL_DEBUG
+  std::unique_ptr<vixl::CodeBufferCheckScope> m_emitter_check;
+  std::unique_ptr<vixl::CodeBufferCheckScope> m_far_emitter_check;
+#endif
+};
+
+} // namespace CPU::NewRec

src/core/cpu_newrec_compiler_aarch64.cpp

@@ -16,7 +16,7 @@
 #include <limits>
 Log_SetChannel(CPU::NewRec);
 
-#define PTR(x) vixl::aarch64::MemOperand(RSTATE, (u32)(((u8*)(x)) - ((u8*)&g_state)))
+#define PTR(x) vixl::aarch64::MemOperand(RSTATE, (((u8*)(x)) - ((u8*)&g_state)))
 
 namespace CPU::NewRec {
 
@@ -1019,10 +1019,10 @@ void CPU::NewRec::AArch64Compiler::Compile_div(CompileFlags cf)
   Label done;
   Label not_divide_by_zero;
   armAsm->cbnz(rt, &not_divide_by_zero);
-  armAsm->cmp(rs, 0);
   armAsm->mov(rhi, rs); // hi = num
   EmitMov(rlo, 1);
   EmitMov(RWSCRATCH, static_cast<u32>(-1));
+  armAsm->cmp(rs, 0);
   armAsm->csel(rlo, RWSCRATCH, rlo, ge); // lo = s >= 0 ? -1 : 1
   armAsm->b(&done);
 
@@ -1328,7 +1328,7 @@ vixl::aarch64::WRegister CPU::NewRec::AArch64Compiler::GenerateLoad(const vixl::
     {
       DebugAssert(addr_reg.GetCode() != RWARG3.GetCode());
       armAsm->lsr(RXARG3, addr_reg, Bus::FASTMEM_LUT_PAGE_SHIFT);
-      armAsm->ldr(RXARG3, MemOperand(RMEMBASE, RXARG3, LSL, 8));
+      armAsm->ldr(RXARG3, MemOperand(RMEMBASE, RXARG3, LSL, 3));
     }
 
     const MemOperand mem =
@@ -1438,7 +1438,7 @@ void CPU::NewRec::AArch64Compiler::GenerateStore(const vixl::aarch64::WRegister&
     {
       DebugAssert(addr_reg.GetCode() != RWARG3.GetCode());
       armAsm->lsr(RXARG3, addr_reg, Bus::FASTMEM_LUT_PAGE_SHIFT);
-      armAsm->ldr(RXARG3, MemOperand(RMEMBASE, RXARG3, LSL, 8));
+      armAsm->ldr(RXARG3, MemOperand(RMEMBASE, RXARG3, LSL, 3));
     }
 
     const MemOperand mem =

src/core/cpu_recompiler_code_generator_aarch32.cpp

@@ -25,8 +25,19 @@ namespace CPU::Recompiler {
 constexpr u32 FUNCTION_CALLEE_SAVED_SPACE_RESERVE = 80;  // 8 registers
 constexpr u32 FUNCTION_CALLER_SAVED_SPACE_RESERVE = 144; // 18 registers -> 224 bytes
 constexpr u32 FUNCTION_STACK_SIZE = FUNCTION_CALLEE_SAVED_SPACE_RESERVE + FUNCTION_CALLER_SAVED_SPACE_RESERVE;
+
+static constexpr u32 TRAMPOLINE_AREA_SIZE = 4 * 1024;
+static std::unordered_map<const void*, u32> s_trampoline_targets;
+static u8* s_trampoline_start_ptr = nullptr;
+static u32 s_trampoline_used = 0;
 } // namespace CPU::Recompiler
 
+bool CPU::Recompiler::armIsCallerSavedRegister(u32 id)
+{
+  return ((id >= 0 && id <= 3) ||  // r0-r3
+          (id == 12 || id == 14)); // sp, pc
+}
+
 s32 CPU::Recompiler::armGetPCDisplacement(const void* current, const void* target)
 {
   Assert(Common::IsAlignedPow2(reinterpret_cast<size_t>(current), 4));
@@ -59,10 +70,19 @@ void CPU::Recompiler::armMoveAddressToReg(vixl::aarch32::Assembler* armAsm, cons
 
 void CPU::Recompiler::armEmitJmp(vixl::aarch32::Assembler* armAsm, const void* ptr, bool force_inline)
 {
-  // TODO: pooling
+  const void* cur = armAsm->GetCursorAddress<const void*>();
+  s32 displacement = armGetPCDisplacement(cur, ptr);
+  bool use_bx = !armIsPCDisplacementInImmediateRange(displacement);
+  if (use_bx && !force_inline)
+  {
+    if (u8* trampoline = armGetJumpTrampoline(ptr); trampoline)
+    {
+      displacement = armGetPCDisplacement(cur, trampoline);
+      use_bx = !armIsPCDisplacementInImmediateRange(displacement);
+    }
+  }
 
-  const s32 displacement = armGetPCDisplacement(armAsm->GetCursorAddress<const void*>(), ptr);
-  if (!armIsPCDisplacementInImmediateRange(displacement))
+  if (use_bx)
   {
     armMoveAddressToReg(armAsm, RSCRATCH, ptr);
     armAsm->bx(RSCRATCH);
@@ -76,10 +96,19 @@ void CPU::Recompiler::armEmitJmp(vixl::aarch32::Assembler* armAsm, const void* p
 
 void CPU::Recompiler::armEmitCall(vixl::aarch32::Assembler* armAsm, const void* ptr, bool force_inline)
 {
-  // TODO: pooling
+  const void* cur = armAsm->GetCursorAddress<const void*>();
+  s32 displacement = armGetPCDisplacement(cur, ptr);
+  bool use_blx = !armIsPCDisplacementInImmediateRange(displacement);
+  if (use_blx && !force_inline)
+  {
+    if (u8* trampoline = armGetJumpTrampoline(ptr); trampoline)
+    {
+      displacement = armGetPCDisplacement(cur, trampoline);
+      use_blx = !armIsPCDisplacementInImmediateRange(displacement);
+    }
+  }
 
-  const s32 displacement = armGetPCDisplacement(armAsm->GetCursorAddress<const void*>(), ptr);
-  if (!armIsPCDisplacementInImmediateRange(displacement))
+  if (use_blx)
   {
     armMoveAddressToReg(armAsm, RSCRATCH, ptr);
     armAsm->blx(RSCRATCH);
@@ -91,6 +120,21 @@ void CPU::Recompiler::armEmitCall(vixl::aarch32::Assembler* armAsm, const void*
   }
 }
 
+void CPU::Recompiler::armEmitCondBranch(vixl::aarch32::Assembler* armAsm, vixl::aarch32::Condition cond, const void* ptr)
+{
+  const s32 displacement = armGetPCDisplacement(armAsm->GetCursorAddress<const void*>(), ptr);
+  if (!armIsPCDisplacementInImmediateRange(displacement))
+  {
+    armMoveAddressToReg(armAsm, RSCRATCH, ptr);
+    armAsm->blx(cond, RSCRATCH);
+  }
+  else
+  {
+    a32::Label label(displacement + armAsm->GetCursorOffset());
+    armAsm->b(cond, &label);
+  }
+}
+
 void CPU::CodeCache::DisassembleAndLogHostCode(const void* start, u32 size)
 {
 #ifdef ENABLE_HOST_DISASSEMBLY
@@ -128,6 +172,36 @@ u32 CPU::CodeCache::EmitJump(void* code, const void* dst, bool flush_icache)
   return kA32InstructionSizeInBytes;
 }
 
+u8* CPU::Recompiler::armGetJumpTrampoline(const void* target)
+{
+  auto it = s_trampoline_targets.find(target);
+  if (it != s_trampoline_targets.end())
+    return s_trampoline_start_ptr + it->second;
+
+  // align to 16 bytes?
+  const u32 offset = s_trampoline_used; // Common::AlignUpPow2(s_trampoline_used, 16);
+
+  // 4 movs plus a jump
+  if (TRAMPOLINE_AREA_SIZE - offset < 20)
+  {
+    Panic("Ran out of space in constant pool");
+    return nullptr;
+  }
+
+  u8* start = s_trampoline_start_ptr + offset;
+  a32::Assembler armAsm(start, TRAMPOLINE_AREA_SIZE - offset);
+  armMoveAddressToReg(&armAsm, RSCRATCH, target);
+  armAsm.bx(RSCRATCH);
+
+  const u32 size = static_cast<u32>(armAsm.GetSizeOfCodeGenerated());
+  DebugAssert(size < 20);
+  s_trampoline_targets.emplace(target, offset);
+  s_trampoline_used = offset + static_cast<u32>(size);
+
+  JitCodeBuffer::FlushInstructionCache(start, size);
+  return start;
+}
+
 u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
 {
   using namespace vixl::aarch32;

src/core/cpu_recompiler_types.h

@@ -78,6 +78,7 @@ constexpr u32 MAX_NEAR_HOST_BYTES_PER_INSTRUCTION = 64;
 constexpr u32 MAX_FAR_HOST_BYTES_PER_INSTRUCTION = 128;
 
 #define RRET vixl::aarch32::r0
+#define RRETHI vixl::aarch32::r1
 #define RARG1 vixl::aarch32::r0
 #define RARG2 vixl::aarch32::r1
 #define RARG3 vixl::aarch32::r2
@@ -85,12 +86,15 @@ constexpr u32 MAX_FAR_HOST_BYTES_PER_INSTRUCTION = 128;
 #define RSCRATCH vixl::aarch32::r12
 #define RSTATE vixl::aarch32::r4
 
+bool armIsCallerSavedRegister(u32 id);
 s32 armGetPCDisplacement(const void* current, const void* target);
 bool armIsPCDisplacementInImmediateRange(s32 displacement);
 void armMoveAddressToReg(vixl::aarch32::Assembler* armAsm, const vixl::aarch32::Register& reg, const void* addr);
 void armEmitMov(vixl::aarch32::Assembler* armAsm, const vixl::aarch32::Register& rd, u32 imm);
 void armEmitJmp(vixl::aarch32::Assembler* armAsm, const void* ptr, bool force_inline);
 void armEmitCall(vixl::aarch32::Assembler* armAsm, const void* ptr, bool force_inline);
+void armEmitCondBranch(vixl::aarch32::Assembler* armAsm, vixl::aarch32::Condition cond, const void* ptr);
+u8* armGetJumpTrampoline(const void* target);
 
 } // namespace CPU::Recompiler