Docs: Update for release FEX-2110

Thunks: Support versioned libraries

.gitmodules

@@ -39,3 +39,10 @@
 [submodule "External/drm-headers"]
 	path = External/drm-headers
 	url = https://github.com/FEX-Emu/drm-headers.git
+[submodule "External/xxhash"]
+	path = External/xxhash
+	url = https://github.com/FEX-Emu/xxHash.git
+[submodule "External/Vulkan-Docs"]
+  shallow = true
+	path = External/Vulkan-Docs
+	url = https://github.com/KhronosGroup/Vulkan-Docs.git

CMakeLists.txt

@@ -14,6 +14,9 @@ option(ENABLE_ASSERTIONS "Enables assertions in build" FALSE)
 option(ENABLE_VISUAL_DEBUGGER "Enables the visual debugger for compiling" FALSE)
 option(ENABLE_STRICT_WERROR "Enables stricter -Werror for CI" FALSE)
 option(ENABLE_WERROR "Enables -Werror" FALSE)
+option(ENABLE_STATIC_PIE "Enables static-pie build" FALSE)
+option(ENABLE_JEMALLOC "Enables jemalloc allocator" TRUE)
+option(ENABLE_OFFLINE_TELEMETRY "Enables FEX offline telemetry" TRUE)
 
 set (X86_C_COMPILER "x86_64-linux-gnu-gcc" CACHE STRING "c compiler for compiling x86 guest libs")
 set (X86_CXX_COMPILER "x86_64-linux-gnu-g++" CACHE STRING "c++ compiler for compiling x86 guest libs")
@@ -44,38 +47,6 @@ else()
   set(CMAKE_INTERPROCEDURAL_OPTIMIZATION FALSE)
 endif()
 
-find_program(CCACHE_PROGRAM ccache)
-if(CCACHE_PROGRAM)
-  message(STATUS "CCache enabled")
-  set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${CCACHE_PROGRAM}")
-endif()
-
-if (ENABLE_XRAY)
-  add_compile_options(-fxray-instrument)
-  link_libraries(-fxray-instrument)
-endif()
-
-if (ENABLE_LLD)
-  link_libraries(-fuse-ld=lld)
-endif()
-
-if (ENABLE_ASAN)
-  add_definitions(-DENABLE_ASAN=1)
-  add_compile_options(-fno-omit-frame-pointer -fsanitize=address -fsanitize-address-use-after-scope)
-  link_libraries(-fno-omit-frame-pointer -fsanitize=address -fsanitize-address-use-after-scope)
-endif()
-
-if (ENABLE_TSAN)
-  add_compile_options(-fno-omit-frame-pointer -fsanitize=thread)
-  link_libraries(-fno-omit-frame-pointer -fsanitize=thread)
-endif()
-
-set (CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} -fno-omit-frame-pointer")
-set (CMAKE_LINKER_FLAGS_RELWITHDEBINFO "${CMAKE_LINKER_FLAGS_RELWITHDEBINFO} -fno-omit-frame-pointer")
-
-set (CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -fomit-frame-pointer")
-set (CMAKE_LINKER_FLAGS_RELEASE "${CMAKE_LINKER_FLAGS_RELEASE} -fomit-frame-pointer")
-
 if (CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64")
   option(ENABLE_X86_HOST_DEBUG "Enables compiling on x86_64 host" FALSE)
   if (NOT ENABLE_X86_HOST_DEBUG)
@@ -95,6 +66,181 @@ if (CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64")
   add_definitions(-D_M_ARM_64=1)
 endif()
 
+find_program(CCACHE_PROGRAM ccache)
+if(CCACHE_PROGRAM)
+  message(STATUS "CCache enabled")
+  set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${CCACHE_PROGRAM}")
+endif()
+
+if (ENABLE_XRAY)
+  add_compile_options(-fxray-instrument)
+  link_libraries(-fxray-instrument)
+endif()
+
+set (PTHREAD_LIB pthread)
+if (ENABLE_LLD)
+  set (LD_OVERRIDE "-fuse-ld=lld")
+  link_libraries(${LD_OVERRIDE})
+endif()
+
+if (NOT ENABLE_OFFLINE_TELEMETRY)
+  # Disable FEX offline telemetry entirely if asked
+  add_definitions(-DFEX_DISABLE_TELEMETRY=1)
+endif()
+
+if (ENABLE_STATIC_PIE)
+  if (_M_ARM_64 AND ENABLE_LLD)
+    message (FATAL_ERROR "Static linking does not currently work with AArch64+LLD. Use GNU ld for now.")
+  endif()
+
+  file(WRITE ${PROJECT_BINARY_DIR}/CMakeFiles/CMakeTmp/Determine_iplt.c
+    "int main(int argc, char* argv[])
+      {
+        return 0;
+      }")
+
+  # Compile the test application with our LD_OVERRIDE and static-pie options
+  try_compile(
+    COMPILE_RESULT
+    ${PROJECT_BINARY_DIR}/CMakeFiles/CMakeTmp
+    ${PROJECT_BINARY_DIR}/CMakeFiles/CMakeTmp/Determine_iplt.c
+    COMPILE_DEFINITIONS "-fPIE ${LD_OVERRIDE}"
+    LINK_LIBRARIES "-static-pie ${LD_OVERRIDE}"
+    COPY_FILE ${PROJECT_BINARY_DIR}/CMakeFiles/CMakeTmp/Determine_iplt
+  )
+
+  if (${COMPILE_RESULT})
+    # Read the symbols from the elf
+    execute_process(COMMAND
+      readelf -s ${PROJECT_BINARY_DIR}/CMakeFiles/CMakeTmp/Determine_iplt
+      OUTPUT_FILE ${PROJECT_BINARY_DIR}/CMakeFiles/CMakeTmp/plt_out.txt
+      OUTPUT_VARIABLE PLT_SYMBOLS)
+
+    # Pull out the __rela_iplt_{start,end} symbols if they exist
+    execute_process(COMMAND
+      "grep" "__rela_iplt" ${PROJECT_BINARY_DIR}/CMakeFiles/CMakeTmp/plt_out.txt
+      OUTPUT_VARIABLE PLT_SYMBOLS)
+
+    set (SYMBOLS_FINE TRUE)
+    set (HAS_IPLT -1)
+    # Check if we have any symbols in our grep output
+    # The symbols must either not exist at all OR the symbols are zero
+    if (PLT_SYMBOLS)
+      string(FIND ${PLT_SYMBOLS} "__rela_iplt_start" HAS_IPLT)
+    endif()
+
+    if (NOT HAS_IPLT EQUAL -1)
+      # We have some symbols from readelf. Let's parse the results to check if they are zero
+      # Format: '35: 0000000000000000     0 NOTYPE  LOCAL  HIDDEN   UND __rela_iplt_start'
+      string(REPLACE "\n" ";" SYMBOL_LIST ${PLT_SYMBOLS})
+      foreach (SYMBOL ${SYMBOL_LIST})
+        # strip any leading and trailing whitespace
+        string (STRIP ${SYMBOL} SYMBOL)
+        # Convert string to a list
+        string(REPLACE " " ";" SYMBOL_VALUES ${SYMBOL}})
+        # Pull out the address argument
+        list(GET SYMBOL_VALUES 1 OFFSET)
+
+        # Check against integer zero
+        if (NOT ${OFFSET} EQUAL 0)
+          # Symbol wasn't zero, this now fails
+          set (SYMBOLS_FINE FALSE)
+        endif()
+      endforeach()
+    endif()
+
+    if (SYMBOLS_FINE)
+      # We can now exnable static-pie
+      set (STATIC_PIE_OPTIONS "-static-pie")
+      # Pthreads has an issue with exposing symbols
+      # We need to make some concessions to the pthread gods
+      if (ENABLE_LLD)
+        set (PTHREAD_LIB
+          -Wl,--undefined-glob=pthread_*
+          -Wl,--undefined=__cxa_finalize
+          -Wl,--undefined=_pthread_cleanup_push_defer
+          -Wl,--undefined=_pthread_cleanup_pop_restore
+          -Wl,--undefined=__pthread_cleanup_upto
+          pthread)
+      else()
+        set (PTHREAD_LIB
+          -Wl,--undefined=pthread_join
+          -Wl,--undefined=pthread_attr_getdetachstate
+          -Wl,--undefined=pthread_sigmask
+          -Wl,--undefined=pthread_mutex_lock
+          -Wl,--undefined=pthread_cond_init
+          -Wl,--undefined=pthread_attr_init
+          -Wl,--undefined=pthread_mutex_unlock
+          -Wl,--undefined=pthread_mutexattr_destroy
+          -Wl,--undefined=pthread_detach
+          -Wl,--undefined=pthread_mutex_init
+          -Wl,--undefined=pthread_getattr_np
+          -Wl,--undefined=pthread_cond_timedwait
+          -Wl,--undefined=pthread_attr_destroy
+          -Wl,--undefined=pthread_mutexattr_settype
+          -Wl,--undefined=pthread_rwlock_unlock
+          -Wl,--undefined=pthread_rwlock_wrlock
+          -Wl,--undefined=pthread_setspecific
+          -Wl,--undefined=pthread_create
+          -Wl,--undefined=pthread_cond_clockwait
+          -Wl,--undefined=pthread_key_create
+          -Wl,--undefined=pthread_rwlock_rdlock
+          -Wl,--undefined=pthread_setname_np
+          -Wl,--undefined=pthread_cond_signal
+          -Wl,--undefined=pthread_mutexattr_init
+          -Wl,--undefined=pthread_attr_setstack
+          -Wl,--undefined=pthread_self
+          -Wl,--undefined=pthread_getaffinity_np
+          -Wl,--undefined=pthread_cond_wait
+          -Wl,--undefined=pthread_mutex_trylock
+          -Wl,--undefined=pthread_cond_broadcast
+          -Wl,--undefined=pthread_cond_destroy
+          -Wl,--undefined=pthread_getspecific
+          -Wl,--undefined=pthread_key_delete
+          -Wl,--undefined=pthread_once
+          -Wl,--undefined=__cxa_finalize
+          -Wl,--undefined=_pthread_cleanup_push_defer
+          -Wl,--undefined=_pthread_cleanup_pop_restore
+          -Wl,--undefined=__pthread_cleanup_upto
+          pthread)
+      endif()
+    else()
+      message (FATAL_ERROR "Application has __rela_iplt_{start,end} symbols. Which means static-pie can't be enabled")
+    endif()
+  else()
+    message (FATAL_ERROR "Couldn't compile static-pie test. Static-pie can't be enabled!")
+  endif()
+endif()
+
+if (ENABLE_ASAN)
+  add_definitions(-DENABLE_ASAN=1)
+  add_compile_options(-fno-omit-frame-pointer -fsanitize=address -fsanitize-address-use-after-scope)
+  link_libraries(-fno-omit-frame-pointer -fsanitize=address -fsanitize-address-use-after-scope)
+endif()
+
+if (ENABLE_TSAN)
+  add_compile_options(-fno-omit-frame-pointer -fsanitize=thread)
+  link_libraries(-fno-omit-frame-pointer -fsanitize=thread)
+endif()
+
+if (ENABLE_JEMALLOC)
+  add_definitions(-DENABLE_JEMALLOC=1)
+  add_subdirectory(External/jemalloc/)
+  include_directories(External/jemalloc/pregen/include/)
+else()
+  message (STATUS
+    " jemalloc disabled!\n"
+    " This is not a recommended configuration!\n"
+    " This will very explicitly break 32-bit application execution!\n"
+    " Use at your own risk!")
+endif()
+
+set (CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} -fno-omit-frame-pointer")
+set (CMAKE_LINKER_FLAGS_RELWITHDEBINFO "${CMAKE_LINKER_FLAGS_RELWITHDEBINFO} -fno-omit-frame-pointer")
+
+set (CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -fomit-frame-pointer")
+set (CMAKE_LINKER_FLAGS_RELEASE "${CMAKE_LINKER_FLAGS_RELEASE} -fomit-frame-pointer")
+
 add_subdirectory(External/vixl/)
 include_directories(External/vixl/src/)
 
@@ -105,14 +251,17 @@ endif()
 
 find_package(PkgConfig REQUIRED)
 find_package(Python 3.0 REQUIRED COMPONENTS Interpreter)
-pkg_check_modules(XXHASH libxxhash REQUIRED)
+pkg_check_modules(XXHASH libxxhash>=0.8.0 QUIET)
+
+if (NOT XXHASH_FOUND)
+  message(STATUS "xxHash not found. Using Externals")
+  add_subdirectory(External/xxhash/)
+  include_directories(External/xxhash/)
+endif()
 
 add_definitions(-Wno-trigraphs)
 add_definitions(-DGLOBAL_DATA_DIRECTORY="${DATA_DIRECTORY}/")
 
-add_subdirectory(External/jemalloc/)
-include_directories(External/jemalloc/pregen/include/)
-
 add_subdirectory(External/cpp-optparse/)
 include_directories(External/cpp-optparse/)
 
@@ -267,6 +416,11 @@ add_subdirectory(Data/binfmts/)
 add_subdirectory(Source/)
 add_subdirectory(Data/AppConfig/)
 
+# Install the ThunksDB file
+install(
+  FILES ${CMAKE_CURRENT_SOURCE_DIR}/Data/ThunksDB.json
+  DESTINATION ${DATA_DIRECTORY}/)
+
 if (BUILD_TESTS)
   add_subdirectory(unittests/)
 endif()
@@ -278,7 +432,10 @@ if (BUILD_THUNKS)
     PREFIX host-libs
     SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/ThunkLibs/HostLibs"
     BINARY_DIR "Host"
-    CMAKE_ARGS "-DCMAKE_INSTALL_PREFIX=${CMAKE_INSTALL_PREFIX}"
+    CMAKE_ARGS
+      "-DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}"
+      "-DCMAKE_INSTALL_PREFIX=${CMAKE_INSTALL_PREFIX}"
+      "-DVULKAN_XML=${CMAKE_SOURCE_DIR}/External/Vulkan-Docs/xml/vk.xml"
     INSTALL_COMMAND ""
     BUILD_ALWAYS ON
   )
@@ -296,7 +453,13 @@ if (BUILD_THUNKS)
     PREFIX guest-libs
     SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/ThunkLibs/GuestLibs"
     BINARY_DIR "Guest"
-    CMAKE_ARGS "-DX86_C_COMPILER:STRING=${X86_C_COMPILER}" "-DX86_CXX_COMPILER:STRING=${X86_CXX_COMPILER}" "-DCMAKE_INSTALL_PREFIX=${CMAKE_INSTALL_PREFIX}"
+    CMAKE_ARGS
+      "-DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}"
+      "-DX86_C_COMPILER:STRING=${X86_C_COMPILER}"
+      "-DX86_CXX_COMPILER:STRING=${X86_CXX_COMPILER}"
+      "-DCMAKE_INSTALL_PREFIX=${CMAKE_INSTALL_PREFIX}"
+      "-DSTRUCT_VERIFIER=${CMAKE_SOURCE_DIR}/Scripts/StructPackVerifier.py"
+      "-DVULKAN_XML=${CMAKE_SOURCE_DIR}/External/Vulkan-Docs/xml/vk.xml"
     INSTALL_COMMAND ""
     BUILD_ALWAYS ON
   )
@@ -356,6 +519,7 @@ endif()
 
 # Package creation
 set (CPACK_GENERATOR "DEB")
+set (CPACK_PACKAGE_NAME fex-emu)
 set (CPACK_PACKAGE_CONTACT "team@fex-emu.org")
 set (CPACK_PACKAGE_VERSION_MAJOR "${FEX_VERSION_MAJOR}")
 set (CPACK_PACKAGE_VERSION_MINOR "${FEX_VERSION_MINOR}")

Data/AppConfig/steamwebhelper.json

@@ -0,0 +1,5 @@
+{
+  "Config": {
+    "StallProcess": "1"
+  }
+}

Data/ThunksDB.json

@@ -0,0 +1,287 @@
+{
+  "DB": {
+    "GL": {
+      "Library" : "libGL-guest.so",
+      "Depends": [
+        "X11"
+      ],
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libGL.so",
+        "/usr/lib/x86_64-linux-gnu/libGL.so.1",
+        "/usr/lib/x86_64-linux-gnu/libGL.so.1.2.0",
+        "/usr/lib/x86_64-linux-gnu/libGL.so.1.7.0",
+        "/lib/x86_64-linux-gnu/libGL.so",
+        "/lib/x86_64-linux-gnu/libGL.so.1",
+        "/lib/x86_64-linux-gnu/libGL.so.1.2.0",
+        "/lib/x86_64-linux-gnu/libGL.so.1.7.0"
+      ]
+    },
+    "GLESv2": {
+      "Library": "libGLESv2-guest.so",
+      "Depends": [
+        "X11"
+      ],
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libGLESv2.so",
+        "/usr/lib/x86_64-linux-gnu/libGLESv2.so.2",
+        "/usr/lib/x86_64-linux-gnu/libGLESv2.so.2.0.0",
+        "/lib/x86_64-linux-gnu/libGLESv2.so",
+        "/lib/x86_64-linux-gnu/libGLESv2.so.2",
+        "/lib/x86_64-linux-gnu/libGLESv2.so.2.0.0"
+      ]
+    },
+    "X11": {
+      "Library": "libX11-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libX11.so",
+        "/usr/lib/x86_64-linux-gnu/libX11.so.6",
+        "/usr/lib/x86_64-linux-gnu/libX11.so.6.4.0",
+        "/lib/x86_64-linux-gnu/libX11.so",
+        "/lib/x86_64-linux-gnu/libX11.so.6",
+        "/lib/x86_64-linux-gnu/libX11.so.6.4.0"
+      ]
+    },
+    "Vulkan-radeon": {
+      "Library": "libvulkan_radeon-guest.so",
+      "Depends": [
+        "xcb"
+      ],
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libvulkan_radeon.so",
+        "/lib/x86_64-linux-gnu/libvulkan_radeon.so"
+      ],
+      "Comment": [
+        "Vulkan library relies on xcb, otherwise it crashes with jemalloc"
+      ]
+    },
+    "Vulkan-lavapipe": {
+      "Library": "libvulkan_lvp-guest.so",
+      "Depends": [
+        "xcb"
+      ],
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libvulkan_lvp.so",
+        "/lib/x86_64-linux-gnu/libvulkan_lvp.so"
+      ]
+    },
+    "Vulkan-freedreno": {
+      "Library": "libvulkan_freedreno-guest.so",
+      "Depends": [
+        "xcb"
+      ],
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libvulkan_freedreno.so",
+        "/lib/x86_64-linux-gnu/libvulkan_freedreno.so"
+      ]
+    },
+    "Vulkan-intel": {
+      "Library": "libvulkan_intel-guest.so",
+      "Depends": [
+        "xcb"
+      ],
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libvulkan_intel.so",
+        "/lib/x86_64-linux-gnu/libvulkan_intel.so"
+      ]
+    },
+    "Vulkan-panfrost": {
+      "Library": "libvulkan_panfrost-guest.so",
+      "Depends": [
+        "xcb"
+      ],
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libvulkan_panfrost.so",
+        "/lib/x86_64-linux-gnu/libvulkan_panfrost.so"
+      ]
+    },
+    "Vulkan-nvidia": {
+      "Library": "libvulkan_nvidia-guest.so",
+      "Depends": [
+        "xcb"
+      ],
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libGLX_nvidia.so.0",
+        "/lib/x86_64-linux-gnu/libGLX_nvidia.so.0"
+      ],
+      "Comment": [
+        "Not currently wired up"
+      ]
+    },
+    "Vulkan-virtio": {
+      "Library": "libvulkan_virtio-guest.so",
+      "Depends": [
+        "xcb"
+      ],
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libvulkan_virtio.so",
+        "/lib/x86_64-linux-gnu/libvulkan_virtio.so"
+      ]
+    },
+    "xcb": {
+      "Library": "libxcb-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libxcb.so",
+        "/usr/lib/x86_64-linux-gnu/libxcb.so.1",
+        "/usr/lib/x86_64-linux-gnu/libxcb.so.1.1.0",
+        "/lib/x86_64-linux-gnu/libxcb.so",
+        "/lib/x86_64-linux-gnu/libxcb.so.1",
+        "/lib/x86_64-linux-gnu/libxcb.so.1.1.0"
+      ]
+    },
+    "xcb-dri2": {
+      "Library": "libxcb_dri2-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libxcb-dri2.so",
+        "/usr/lib/x86_64-linux-gnu/libxcb-dri2.so.0",
+        "/usr/lib/x86_64-linux-gnu/libxcb-dri2.so.0.0.0",
+        "/lib/x86_64-linux-gnu/libxcb-dri2.so",
+        "/lib/x86_64-linux-gnu/libxcb-dri2.so.0",
+        "/lib/x86_64-linux-gnu/libxcb-dri2.so.0.0.0"
+      ]
+    },
+    "xcb-dri3": {
+      "Library": "libxcb_dri3-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libxcb-dri3.so",
+        "/usr/lib/x86_64-linux-gnu/libxcb-dri3.so.0",
+        "/usr/lib/x86_64-linux-gnu/libxcb-dri3.so.0.0.0",
+        "/lib/x86_64-linux-gnu/libxcb-dri3.so",
+        "/lib/x86_64-linux-gnu/libxcb-dri3.so.0",
+        "/lib/x86_64-linux-gnu/libxcb-dri3.so.0.0.0"
+      ]
+    },
+    "xcb-xfixes": {
+      "Library": "libxcb_xfixes-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libxcb-xfixes.so",
+        "/usr/lib/x86_64-linux-gnu/libxcb-xfixes.so.0",
+        "/usr/lib/x86_64-linux-gnu/libxcb-xfixes.so.0.0.0",
+        "/lib/x86_64-linux-gnu/libxcb-xfixes.so",
+        "/lib/x86_64-linux-gnu/libxcb-xfixes.so.0",
+        "/lib/x86_64-linux-gnu/libxcb-xfixes.so.0.0.0"
+      ]
+    },
+    "xcb-shm": {
+      "Library": "libxcb_shm-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libxcb-shm.so",
+        "/usr/lib/x86_64-linux-gnu/libxcb-shm.so.0",
+        "/usr/lib/x86_64-linux-gnu/libxcb-shm.so.0.0.0",
+        "/lib/x86_64-linux-gnu/libxcb-shm.so",
+        "/lib/x86_64-linux-gnu/libxcb-shm.so.0",
+        "/lib/x86_64-linux-gnu/libxcb-shm.so.0.0.0"
+      ]
+    },
+    "xcb-sync": {
+      "Library": "libxcb_sync-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libxcb-sync.so",
+        "/usr/lib/x86_64-linux-gnu/libxcb-sync.so.1",
+        "/usr/lib/x86_64-linux-gnu/libxcb-sync.so.1.0.0",
+        "/lib/x86_64-linux-gnu/libxcb-sync.so",
+        "/lib/x86_64-linux-gnu/libxcb-sync.so.1",
+        "/lib/x86_64-linux-gnu/libxcb-sync.so.1.0.0"
+      ]
+    },
+    "xcb-randr": {
+      "Library": "libxcb_randr-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libxcb-randr.so",
+        "/usr/lib/x86_64-linux-gnu/libxcb-randr.so.0",
+        "/usr/lib/x86_64-linux-gnu/libxcb-randr.so.0.1.0",
+        "/lib/x86_64-linux-gnu/libxcb-randr.so",
+        "/lib/x86_64-linux-gnu/libxcb-randr.so.0",
+        "/lib/x86_64-linux-gnu/libxcb-randr.so.0.1.0"
+      ]
+    },
+    "xcb-present": {
+      "Library": "libxcb_present-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libxcb-present.so",
+        "/usr/lib/x86_64-linux-gnu/libxcb-present.so.0",
+        "/usr/lib/x86_64-linux-gnu/libxcb-present.so.0.0.0",
+        "/lib/x86_64-linux-gnu/libxcb-present.so",
+        "/lib/x86_64-linux-gnu/libxcb-present.so.0",
+        "/lib/x86_64-linux-gnu/libxcb-present.so.0.0.0"
+      ]
+    },
+    "xcb-glx": {
+      "Library": "libxcb_glx-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libxcb-glx.so",
+        "/usr/lib/x86_64-linux-gnu/libxcb-glx.so.0",
+        "/usr/lib/x86_64-linux-gnu/libxcb-glx.so.0.0.0",
+        "/lib/x86_64-linux-gnu/libxcb-glx.so",
+        "/lib/x86_64-linux-gnu/libxcb-glx.so.0",
+        "/lib/x86_64-linux-gnu/libxcb-glx.so.0.0.0"
+      ]
+    },
+    "xshmfence": {
+      "Library": "libshmfence-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libxshmfence.so",
+        "/usr/lib/x86_64-linux-gnu/libxshmfence.so.1",
+        "/usr/lib/x86_64-linux-gnu/libxshmfence.so.1.0.0",
+        "/lib/x86_64-linux-gnu/libxshmfence.so",
+        "/lib/x86_64-linux-gnu/libxshmfence.so.1",
+        "/lib/x86_64-linux-gnu/libxshmfence.so.1.0.0"
+      ]
+    },
+    "drm": {
+      "Library": "libdrm-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libdrm.so",
+        "/usr/lib/x86_64-linux-gnu/libdrm.so.2",
+        "/usr/lib/x86_64-linux-gnu/libdrm.so.2.4.0",
+        "/lib/x86_64-linux-gnu/libdrm.so",
+        "/lib/x86_64-linux-gnu/libdrm.so.2",
+        "/lib/x86_64-linux-gnu/libdrm.so.2.4.0"
+      ]
+    },
+    "asound": {
+      "Library": "libasound-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libasound.so",
+        "/usr/lib/x86_64-linux-gnu/libasound.so.2",
+        "/usr/lib/x86_64-linux-gnu/libasound.so.2.0.0",
+        "/lib/x86_64-linux-gnu/libasound.so",
+        "/lib/x86_64-linux-gnu/libasound.so.2",
+        "/lib/x86_64-linux-gnu/libasound.so.2.0.0"
+      ]
+    },
+    "Xrender": {
+      "Library": "libXrender-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libXrender.so",
+        "/usr/lib/x86_64-linux-gnu/libXrender.so.1",
+        "/usr/lib/x86_64-linux-gnu/libXrender.so.1.3.0",
+        "/lib/x86_64-linux-gnu/libXrender.so",
+        "/lib/x86_64-linux-gnu/libXrender.so.1",
+        "/lib/x86_64-linux-gnu/libXrender.so.1.3.0"
+      ]
+    },
+    "Xext": {
+      "Library": "libXext-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libXext.so",
+        "/usr/lib/x86_64-linux-gnu/libXext.so.6",
+        "/usr/lib/x86_64-linux-gnu/libXext.so.6.4.0",
+        "/lib/x86_64-linux-gnu/libXext.so",
+        "/lib/x86_64-linux-gnu/libXext.so.6",
+        "/lib/x86_64-linux-gnu/libXext.so.6.4.0"
+      ]
+    },
+    "Xfixes": {
+      "Library": "libXfixes-guest.so",
+      "Overlay": [
+        "/usr/lib/x86_64-linux-gnu/libXfixes.so",
+        "/usr/lib/x86_64-linux-gnu/libXfixes.so.3",
+        "/usr/lib/x86_64-linux-gnu/libXfixes.so.3.1.0",
+        "/lib/x86_64-linux-gnu/libXfixes.so",
+        "/lib/x86_64-linux-gnu/libXfixes.so.3",
+        "/lib/x86_64-linux-gnu/libXfixes.so.3.1.0"
+      ]
+    },
+    "":{}
+  }
+}

Data/binfmts/CMakeLists.txt

@@ -1,4 +1,17 @@
-install(FILES FEX-x86
-  DESTINATION ${CMAKE_INSTALL_PREFIX}/share/binfmts/)
-install(FILES FEX-x86_64
-  DESTINATION ${CMAKE_INSTALL_PREFIX}/share/binfmts/)
+function(GenBinFmt Name)
+  # Get the filename only component
+  get_filename_component(FMT_NAME ${Name} NAME_WE)
+
+  # Configure it
+  configure_file(
+    ${Name}
+    ${CMAKE_BINARY_DIR}/Data/binfmts/${FMT_NAME})
+
+  # Then install the configured binfmt
+  install(
+    FILES ${CMAKE_BINARY_DIR}/Data/binfmts/${FMT_NAME}
+    DESTINATION ${CMAKE_INSTALL_PREFIX}/share/binfmts/)
+endfunction()
+
+GenBinFmt(FEX-x86.in)
+GenBinFmt(FEX-x86_64.in)

Data/binfmts/FEX-x86.in

@@ -1,9 +1,8 @@
 package fex
-interpreter /usr/bin/FEXInterpreter
+interpreter @CMAKE_INSTALL_PREFIX@/bin/FEXInterpreter
 magic \x7fELF\x01\x01\x01\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x03\x00
 offset 0
 mask \xff\xff\xff\xff\xff\xfe\xfe\x00\x00\x00\x00\xff\xff\xff\xff\xff\xfe\xff\xff\xff
 credentials yes
 fix_binary yes
-preserve no
-
+preserve yes

Data/binfmts/FEX-x86_64.in

@@ -1,8 +1,8 @@
 package fex
-interpreter /usr/bin/FEXInterpreter
+interpreter @CMAKE_INSTALL_PREFIX@/bin/FEXInterpreter
 magic \x7fELF\x02\x01\x01\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x3e\x00
 offset 0
 mask \xff\xff\xff\xff\xff\xfe\xfe\x00\x00\x00\x00\xff\xff\xff\xff\xff\xfe\xff\xff\xff
 credentials yes
 fix_binary yes
-preserve no
+preserve yes

External/FEXCore/Scripts/config_generator.py

@@ -98,14 +98,19 @@ def print_man_option(short, long, desc, default):
     output_man.write("\\fBdefault:\\fR {0}\n".format(default))
     output_man.write(".Pp\n\n")
 
-def print_man_env_option(name, desc, default):
-    output_man.write("\\fBFEX_{0}\\fR\n".format(name))
+def print_man_env_option(name, desc, default, no_json_key):
+    output_man.write("\\fBFEX_{0}\\fR\n".format(name.upper()))
 
     # Print description
     for line in desc:
         output_man.write(".Pp\n")
         output_man.write("{0}\n".format(line))
 
+    if (not no_json_key):
+        output_man.write(".Pp\n")
+        output_man.write("\\fBJSON key:\\fR '{0}'\n".format(name))
+        output_man.write(".Pp\n\n")
+
     output_man.write(".Pp\n")
     output_man.write("\\fBdefault:\\fR {0}\n".format(default))
     output_man.write(".Pp\n\n")
@@ -154,9 +159,10 @@ def print_man_environment(options):
                 # Wrap the string argument in quotes
                 default = "'" + default + "'"
             print_man_env_option(
-                op_key.upper(),
+                op_key,
                 op_vals["Desc"],
-                default
+                default,
+                False
             )
 
     print_man_environment_tail()
@@ -172,7 +178,7 @@ def print_man_environment_tail():
     "By default FEX will look in {$HOME, $XDG_CONFIG_HOME}/.fex-emu/",
     "This will override the full path",
     ],
-    "''")
+    "''", True)
 
     print_man_env_option(
     "FEX_APP_CONFIG",
@@ -183,7 +189,7 @@ def print_man_environment_tail():
     "One must be careful with this option as it will override any applications that load with execve as well"
     "If you need to support applications that execve then use FEX_APP_CONFIG_LOCATION instead"
     ],
-    "''")
+    "''", True)
 
     print_man_env_option(
     "FEX_APP_DATA_LOCATION",
@@ -193,7 +199,7 @@ def print_man_environment_tail():
     "This will override the full path",
     "This is the folder where FEX stores generated files like IR cache"
     ],
-    "''")
+    "''", True)
 
 def print_man_header():
     header ='''.Dd {0}

External/FEXCore/Source/CMakeLists.txt

@@ -81,7 +81,12 @@ set (SRCS
   Interface/Core/Frontend.cpp
   Interface/Core/GdbServer.cpp
   Interface/Core/HostFeatures.cpp
+  Interface/Core/OpcodeDispatcher/Crypto.cpp
+  Interface/Core/OpcodeDispatcher/Flags.cpp
+  Interface/Core/OpcodeDispatcher/Vector.cpp
+  Interface/Core/OpcodeDispatcher/X87.cpp
   Interface/Core/OpcodeDispatcher.cpp
+  Interface/Core/SignalDelegator.cpp
   Interface/Core/X86Tables.cpp
   Interface/Core/X86DebugInfo.cpp
   Interface/Core/X86HelperGen.cpp
@@ -114,6 +119,7 @@ set (SRCS
   Interface/IR/Passes/DeadContextStoreElimination.cpp
   Interface/IR/Passes/IRCompaction.cpp
   Interface/IR/Passes/IRValidation.cpp
+  Interface/IR/Passes/RAValidation.cpp
   Interface/IR/Passes/LongDivideRemovalPass.cpp
   Interface/IR/Passes/ValueDominanceValidation.cpp
   Interface/IR/Passes/PhiValidation.cpp
@@ -125,6 +131,7 @@ set (SRCS
   Utils/Allocator.cpp
   Utils/Allocator/64BitAllocator.cpp
   Utils/LogManager.cpp
+  Utils/Telemetry.cpp
   Utils/Threads.cpp
   )
 
@@ -133,7 +140,7 @@ if(_M_ARM_64)
     Interface/Core/ArchHelpers/Arm64.cpp)
 endif()
 
-set(DEFINES )
+set(DEFINES -DTHREAD_LOCAL=_Thread_local)
 
 if (_M_X86_64)
   list(APPEND DEFINES -D_M_X86_64=1)
@@ -179,6 +186,16 @@ if (ENABLE_JITSYMBOLS)
   list(APPEND DEFINES -DENABLE_JITSYMBOLS=1)
 endif()
 
+set (LIBS vixl dl fmt::fmt xxhash tiny-json)
+if (ENABLE_JEMALLOC)
+  list (APPEND LIBS FEX_jemalloc)
+endif()
+
+# Generate config
+configure_file(
+  ${CMAKE_CURRENT_SOURCE_DIR}/Interface/Config/Config.json.in
+  ${CMAKE_BINARY_DIR}/generated/Config/Config.json)
+
 # Generate IR include file
 set(OUTPUT_IR_FOLDER "${CMAKE_BINARY_DIR}/include/FEXCore/IR")
 set(OUTPUT_NAME "${OUTPUT_IR_FOLDER}/IRDefines.inc")
@@ -221,7 +238,7 @@ add_custom_target(IR_INC
 set(OUTPUT_CONFIG_FOLDER "${CMAKE_BINARY_DIR}/include/FEXCore/Config")
 set(OUTPUT_CONFIG_NAME "${OUTPUT_CONFIG_FOLDER}/ConfigValues.inl")
 set(OUTPUT_CONFIG_OPTION_NAME "${OUTPUT_CONFIG_FOLDER}/ConfigOptions.inl")
-set(INPUT_CONFIG_NAME "${CMAKE_CURRENT_SOURCE_DIR}/Interface/Config/Config.json")
+set(INPUT_CONFIG_NAME "${CMAKE_BINARY_DIR}/generated/Config/Config.json")
 set(OUTPUT_MAN_NAME "${CMAKE_BINARY_DIR}/generated/FEX.1")
 
 add_custom_target(CREATE_CONFIG_FOLDER ALL
@@ -265,7 +282,7 @@ function(AddObject Name Type)
   add_dependencies(${Name} IR_INC)
   add_dependencies(${Name} CONFIG_INC)
 
-  target_link_libraries(${Name} pthread vixl dl fmt::fmt xxhash FEX_jemalloc)
+  target_link_libraries(${Name} ${LIBS})
   set_target_properties(${Name} PROPERTIES OUTPUT_NAME FEXCore)
   set_target_properties(${Name} PROPERTIES C_VISIBILITY_PRESET hidden)
   set_target_properties(${Name} PROPERTIES CXX_VISIBILITY_PRESET hidden)
@@ -306,7 +323,7 @@ endfunction()
 
 function(AddLibrary Name Type)
   add_library(${Name} ${Type} $<TARGET_OBJECTS:${PROJECT_NAME}_object>)
-  target_link_libraries(${Name} pthread vixl dl fmt::fmt xxhash FEX_jemalloc)
+  target_link_libraries(${Name} ${LIBS})
   set_target_properties(${Name} PROPERTIES OUTPUT_NAME FEXCore)
   set_target_properties(${Name} PROPERTIES C_VISIBILITY_PRESET hidden)
   set_target_properties(${Name} PROPERTIES CXX_VISIBILITY_PRESET hidden)

External/FEXCore/Source/Common/NetStream.cpp

@@ -1,11 +1,7 @@
 #include "NetStream.h"
 
-#include <cstring>
-
-#include <sys/types.h>
 #include <sys/socket.h>
 
-#include <stdio.h>
 #include <unistd.h>
 
 int NetStream::NetBuf::flushBuffer(const char *buffer, size_t size) {
@@ -29,7 +25,7 @@ std::streamsize NetStream::NetBuf::xsputn(const char* buffer, std::streamsize si
 
     // Check if the string fits neatly in our buffer
     if (size <= buf_remaining) {
-        std::memcpy(pptr(), buffer, size);
+        ::memcpy(pptr(), buffer, size);
         pbump(size);
         return size;
     }

External/FEXCore/Source/Common/NetStream.h

@@ -2,6 +2,7 @@
 
 #include <array>
 #include <iostream>
+#include <iterator>
 #include <string.h>
 
 class NetStream : public std::iostream {

External/FEXCore/Source/Common/Paths.cpp

@@ -3,12 +3,44 @@
 
 #include <cstdlib>
 #include <filesystem>
-#include <sys/stat.h>
+#include <memory>
+#include <pwd.h>
+#include <system_error>
+#include <unistd.h>
 
 namespace FEXCore::Paths {
   std::unique_ptr<std::string> CachePath;
   std::unique_ptr<std::string> EntryCache;
 
+  char const* FindUserHomeThroughUID() {
+    auto passwd = getpwuid(geteuid());
+    if (passwd) {
+      return passwd->pw_dir;
+    }
+    return nullptr;
+  }
+
+  const char *GetHomeDirectory() {
+    char const *HomeDir = getenv("HOME");
+
+    // Try to get home directory from uid
+    if (!HomeDir) {
+      HomeDir = FindUserHomeThroughUID();
+    }
+
+    // try the PWD
+    if (!HomeDir) {
+      HomeDir = getenv("PWD");
+    }
+
+    // Still doesn't exit? You get local
+    if (!HomeDir) {
+      HomeDir = ".";
+    }
+
+    return HomeDir;
+  }
+
   void InitializePaths() {
     CachePath = std::make_unique<std::string>();
     EntryCache = std::make_unique<std::string>();
@@ -36,9 +68,10 @@ namespace FEXCore::Paths {
     *CachePath += "/.fex-emu/";
     *EntryCache = *CachePath + "/EntryCache/";
 
+    std::error_code ec{};
     // Ensure the folder structure is created for our Data
-    if (!std::filesystem::exists(*EntryCache) &&
-        !std::filesystem::create_directories(*EntryCache)) {
+    if (!std::filesystem::exists(*EntryCache, ec) &&
+        !std::filesystem::create_directories(*EntryCache, ec)) {
       LogMan::Msg::D("Couldn't create EntryCache directory: '%s'", EntryCache->c_str());
     }
   }

External/FEXCore/Source/Common/Paths.h

@@ -4,6 +4,9 @@
 namespace FEXCore::Paths {
   void InitializePaths();
   void ShutdownPaths();
+
+  const char *GetHomeDirectory();
+
   std::string GetCachePath();
   std::string GetEntryCachePath();
 }

External/FEXCore/Source/Common/SoftFloat.h

@@ -16,9 +16,19 @@ extern "C" {
 
 struct X80SoftFloat {
 #ifdef _M_X86_64
+// Define this to push some operations to x87
+// Only useful to see if precision loss is killing something
+// #define DEBUG_X86_FLOAT
+#ifdef DEBUG_X86_FLOAT
+#define BIGFLOAT long double
+#define BIGFLOATSIZE 10
+#else
 #define BIGFLOAT __float128
+#define BIGFLOATSIZE 16
+#endif
 #elif defined(_M_ARM_64)
 #define BIGFLOAT long double
+#define BIGFLOATSIZE 16
 #else
 #error No 128bit float for this target!
 #endif
@@ -170,8 +180,14 @@ struct X80SoftFloat {
   }
 
   operator BIGFLOAT() const {
+#if BIGFLOATSIZE == 16
     const float128_t Result = extF80_to_f128(*this);
     return FEXCore::BitCast<BIGFLOAT>(Result);
+#else
+    BIGFLOAT result{};
+    memcpy(&result, this, sizeof(result));
+    return result;
+#endif
   }
 
   operator int16_t() const {
@@ -217,6 +233,12 @@ struct X80SoftFloat {
     *this = ui64_to_extF80(rhs);
   }
 
+#if BIGFLOATSIZE == 10
+  void operator=(const long double rhs) {
+    memcpy(this, &rhs, sizeof(rhs));
+  }
+#endif
+
   operator void*() {
     return reinterpret_cast<void*>(this);
   }
@@ -236,7 +258,11 @@ struct X80SoftFloat {
   }
 
   X80SoftFloat(BIGFLOAT rhs) {
+#if BIGFLOATSIZE == 16
     *this = f128_to_extF80(FEXCore::BitCast<float128_t>(rhs));
+#else
+    *this = FEXCore::BitCast<long double>(rhs);
+#endif
   }
 
   X80SoftFloat(const int16_t rhs) {

External/FEXCore/Source/Interface/Config/Config.cpp

@@ -1,43 +1,164 @@
 #include "Common/StringConv.h"
-
-#include <FEXCore/Utils/LogManager.h>
-#include "Interface/Context/Context.h"
+#include "Common/Paths.h"
 
 #include <FEXCore/Config/Config.h>
+#include <FEXCore/Utils/LogManager.h>
+
+#include <array>
+#include <assert.h>
+#include <cstdlib>
 #include <filesystem>
-#include <pwd.h>
+#include <fstream>
+#include <functional>
 #include <map>
+#include <memory>
+#include <list>
+#include <optional>
+#include <stddef.h>
+#include <stdint.h>
+#include <string>
+#include <string_view>
 #include <sys/sysinfo.h>
-#include <unistd.h>
+#include <system_error>
+#include <type_traits>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <tiny-json.h>
+
+namespace FEXCore::Context {
+  struct Context;
+}
 
 namespace FEXCore::Config {
-  char const* FindUserHomeThroughUID() {
-    auto passwd = getpwuid(geteuid());
-    if (passwd) {
-      return passwd->pw_dir;
-    }
-    return nullptr;
+namespace DefaultValues {
+#define P(x) x
+#define OPT_BASE(type, group, enum, json, default) const P(type) P(enum) = P(default);
+#define OPT_STR(group, enum, json, default) const std::string_view P(enum) = P(default);
+#define OPT_STRARRAY(group, enum, json, default) OPT_STR(group, enum, json, default)
+#include <FEXCore/Config/ConfigValues.inl>
+}
+
+static bool LoadConfigFile(std::vector<char> &Data, const std::string &Config) {
+  std::fstream ConfigFile;
+  ConfigFile.open(Config, std::ios::in);
+
+  if (!ConfigFile.is_open()) {
+    return false;
   }
 
-  const char *GetHomeDirectory() {
-    char const *HomeDir = getenv("HOME");
+  if (!ConfigFile.seekg(0, std::fstream::end)) {
+    LogMan::Msg::D("Couldn't load configuration file: Seek end");
+    return false;
+  }
 
-    // Try to get home directory from uid
-    if (!HomeDir) {
-      HomeDir = FindUserHomeThroughUID();
+  auto FileSize = ConfigFile.tellg();
+  if (ConfigFile.fail()) {
+    LogMan::Msg::D("Couldn't load configuration file: tellg");
+    return false;
+  }
+
+  if (!ConfigFile.seekg(0, std::fstream::beg)) {
+    LogMan::Msg::D("Couldn't load configuration file: Seek beginning");
+    return false;
+  }
+
+  if (FileSize > 0) {
+    Data.resize(FileSize);
+    if (!ConfigFile.read(&Data.at(0), FileSize)) {
+      // Probably means permissions aren't set. Just early exit
+      return false;
+    }
+    ConfigFile.close();
+  }
+  else {
+    return false;
+  }
+
+  return true;
+}
+
+namespace JSON {
+  struct JsonAllocator {
+    jsonPool_t PoolObject;
+    std::unique_ptr<std::list<json_t>> json_objects;
+  };
+  static_assert(offsetof(JsonAllocator, PoolObject) == 0, "This needs to be at offset zero");
+
+  json_t* PoolInit(jsonPool_t* Pool) {
+    JsonAllocator* alloc = reinterpret_cast<JsonAllocator*>(Pool);
+    alloc->json_objects = std::make_unique<std::list<json_t>>();
+    return &*alloc->json_objects->emplace(alloc->json_objects->end());
+  }
+
+  json_t* PoolAlloc(jsonPool_t* Pool) {
+    JsonAllocator* alloc = reinterpret_cast<JsonAllocator*>(Pool);
+    return &*alloc->json_objects->emplace(alloc->json_objects->end());
+  }
+
+  static void LoadJSonConfig(const std::string &Config, std::function<void(const char *Name, const char *ConfigSring)> Func) {
+    std::vector<char> Data;
+    if (!LoadConfigFile(Data, Config)) {
+      return;
     }
 
-    // try the PWD
-    if (!HomeDir) {
-      HomeDir = getenv("PWD");
+    JsonAllocator Pool {
+      .PoolObject = {
+        .init = PoolInit,
+        .alloc = PoolAlloc,
+      },
+    };
+
+    json_t const *json = json_createWithPool(&Data.at(0), &Pool.PoolObject);
+    if (!json) {
+      LogMan::Msg::E("Couldn't create json");
+      return;
     }
 
-    // Still doesn't exit? You get local
-    if (!HomeDir) {
-      HomeDir = ".";
+    json_t const* ConfigList = json_getProperty(json, "Config");
+
+    if (!ConfigList) {
+      LogMan::Msg::E("Couldn't get config list");
+      return;
     }
 
-    return HomeDir;
+    for (json_t const* ConfigItem = json_getChild(ConfigList);
+      ConfigItem != nullptr;
+      ConfigItem = json_getSibling(ConfigItem)) {
+      const char* ConfigName = json_getName(ConfigItem);
+      const char* ConfigString = json_getValue(ConfigItem);
+
+      if (!ConfigName) {
+        LogMan::Msg::E("Couldn't get config name");
+        return;
+      }
+
+      if (!ConfigString) {
+        LogMan::Msg::E("Couldn't get ConfigString for '%s'", ConfigName);
+        return;
+      }
+
+      Func(ConfigName, ConfigString);
+    }
+  }
+}
+
+  std::string GetDataDirectory() {
+    std::string DataDir{};
+
+    char const *HomeDir = Paths::GetHomeDirectory();
+    char const *DataXDG = getenv("XDG_DATA_HOME");
+    char const *DataOverride = getenv("FEX_APP_DATA_LOCATION");
+    if (DataOverride) {
+      // Data override will override the complete directory
+      DataDir = DataOverride;
+    }
+    else {
+      DataDir = DataXDG ?: HomeDir;
+      DataDir += "/.fex-emu/";
+    }
+    return DataDir;
   }
 
   std::string GetConfigDirectory(bool Global) {
@@ -46,7 +167,7 @@ namespace FEXCore::Config {
       ConfigDir = GLOBAL_DATA_DIRECTORY;
     }
     else {
-      char const *HomeDir = GetHomeDirectory();
+      char const *HomeDir = Paths::GetHomeDirectory();
       char const *ConfigXDG = getenv("XDG_CONFIG_HOME");
       char const *ConfigOverride = getenv("FEX_APP_CONFIG_LOCATION");
       if (ConfigOverride) {
@@ -59,9 +180,9 @@ namespace FEXCore::Config {
       }
 
       // Ensure the folder structure is created for our configuration
-      if (!std::filesystem::exists(ConfigDir) &&
-          !std::filesystem::create_directories(ConfigDir)) {
-        LogMan::Msg::D("Couldn't create config directory: '%s'", ConfigDir.c_str());
+      std::error_code ec{};
+      if (!std::filesystem::exists(ConfigDir, ec) &&
+          !std::filesystem::create_directories(ConfigDir, ec)) {
         // Let's go local in this case
         return "./";
       }
@@ -85,9 +206,11 @@ namespace FEXCore::Config {
 
   std::string GetApplicationConfig(const std::string &Filename, bool Global) {
     std::string ConfigFile = GetConfigDirectory(Global);
+
+    std::error_code ec{};
     if (!Global &&
-        !std::filesystem::exists(ConfigFile) &&
-        !std::filesystem::create_directories(ConfigFile)) {
+        !std::filesystem::exists(ConfigFile, ec) &&
+        !std::filesystem::create_directories(ConfigFile, ec)) {
       LogMan::Msg::D("Couldn't create config directory: '%s'", ConfigFile.c_str());
       // Let's go local in this case
       return "./" + Filename + ".json";
@@ -97,9 +220,8 @@ namespace FEXCore::Config {
 
     // Attempt to create the local folder if it doesn't exist
     if (!Global &&
-        !std::filesystem::exists(ConfigFile) &&
-        !std::filesystem::create_directories(ConfigFile)) {
-      LogMan::Msg::D("Couldn't create AppConfig directory: '%s'", ConfigFile.c_str());
+        !std::filesystem::exists(ConfigFile, ec) &&
+        !std::filesystem::create_directories(ConfigFile, ec)) {
       // Let's go local in this case
       return "./" + Filename + ".json";
     }
@@ -108,23 +230,6 @@ namespace FEXCore::Config {
     return ConfigFile;
   }
 
-  std::string GetDataDirectory() {
-    std::string DataDir{};
-
-    char const *HomeDir = GetHomeDirectory();
-    char const *DataXDG = getenv("XDG_DATA_HOME");
-    char const *DataOverride = getenv("FEX_APP_DATA_LOCATION");
-    if (DataOverride) {
-      // Data override will override the complete directory
-      DataDir = DataOverride;
-    }
-    else {
-      DataDir = DataXDG ?: HomeDir;
-      DataDir += "/.fex-emu/";
-    }
-    return DataDir;
-  }
-
   void SetConfig(FEXCore::Context::Context *CTX, ConfigOption Option, uint64_t Config) {
   }
 
@@ -224,7 +329,8 @@ namespace FEXCore::Config {
   void MetaLayer::MergeConfigMap(const LayerOptions &Options) {
     // Insert this layer's options, overlaying previous options that exist here
     for (auto &it : Options) {
-      if (it.first == FEXCore::Config::ConfigOption::CONFIG_ENV) {
+      if (it.first == FEXCore::Config::ConfigOption::CONFIG_ENV ||
+          it.first == FEXCore::Config::ConfigOption::CONFIG_HOSTENV) {
         MergeEnvironmentVariables(it.first, it.second);
       }
       else {
@@ -252,7 +358,7 @@ namespace FEXCore::Config {
     }
   }
 
-  std::string ExpandPath(std::string PathName) {
+  std::string ExpandPath(std::string const &ContainerPrefix, std::string PathName) {
     if (PathName.empty()) {
       return {};
     }
@@ -273,10 +379,74 @@ namespace FEXCore::Config {
       Path = std::filesystem::absolute(Path);
 
       // Only return if it exists
-      if (std::filesystem::exists(Path)) {
+      std::error_code ec{};
+      if (std::filesystem::exists(Path, ec)) {
         return Path;
       }
     }
+    else {
+      // If the containerprefix and pathname isn't empty
+      // Then we check if the pathname exists in our current namespace
+      // If the path DOESN'T exist but DOES exist with the prefix applied
+      // then redirect to the prefix
+      //
+      // This might not be expected behaviour for some edge cases but since
+      // all paths aren't mounted inside the container, then it'll be fine
+      //
+      // Main catch case for this is the default thunk install folders
+      // HostThunks: $CMAKE_INSTALL_PREFIX/lib/fex-emu/HostThunks/
+      // GuestThunks: $CMAKE_INSTALL_PREFIX/share/fex-emu/GuestThunks/
+      if (!ContainerPrefix.empty() && !PathName.empty()) {
+        if (!std::filesystem::exists(PathName)) {
+          auto ContainerPath = ContainerPrefix + PathName;
+          if (std::filesystem::exists(ContainerPath)) {
+            return ContainerPath;
+          }
+        }
+      }
+    }
+    return {};
+  }
+
+  std::string ltrim(std::string String) {
+    size_t pos = std::string::npos;
+    if ((pos = String.find_first_not_of(" \t\n\r")) != std::string::npos) {
+      String.erase(0, pos);
+    }
+
+    return String;
+  }
+
+  std::string rtrim(std::string String) {
+    size_t pos = std::string::npos;
+    if ((pos = String.find_last_not_of(" \t\n\r")) != std::string::npos) {
+      String.erase(String.begin() + pos + 1, String.end());
+    }
+
+    return String;
+  }
+
+  std::string trim(std::string String) {
+    return rtrim(ltrim(String));
+  }
+
+
+  std::string FindContainerPrefix() {
+    // We only support pressure-vessel at the moment
+    const static std::string ContainerManager = "/run/host/container-manager";
+    if (std::filesystem::exists(ContainerManager)) {
+      std::vector<char> Manager{};
+      if (LoadConfigFile(Manager, ContainerManager)) {
+        // Trim the whitespace, may contain a newline
+        std::string ManagerStr = Manager.data();
+        ManagerStr = trim(ManagerStr);
+        if (strncmp(ManagerStr.data(), "pressure-vessel", Manager.size()) == 0) {
+          // We are running inside of pressure vessel
+          // Our $CMAKE_INSTALL_PREFIX paths are now inside of /run/host/$CMAKE_INSTALL_PREFIX
+          return "/run/host/";
+        }
+      }
+    }
     return {};
   }
 
@@ -292,8 +462,9 @@ namespace FEXCore::Config {
       }
     }
 
-    auto ExpandPathIfExists = [](FEXCore::Config::ConfigOption Config, std::string PathName) {
-      auto NewPath = ExpandPath(PathName);
+    std::string ContainerPrefix { FindContainerPrefix() };
+    auto ExpandPathIfExists = [&ContainerPrefix](FEXCore::Config::ConfigOption Config, std::string PathName) {
+      auto NewPath = ExpandPath(ContainerPrefix, PathName);
       if (!NewPath.empty()) {
         FEXCore::Config::EraseSet(Config, NewPath);
       }
@@ -301,7 +472,7 @@ namespace FEXCore::Config {
 
     if (FEXCore::Config::Exists(FEXCore::Config::CONFIG_ROOTFS)) {
       FEX_CONFIG_OPT(PathName, ROOTFS);
-      auto ExpandedString = ExpandPath(PathName());
+      auto ExpandedString = ExpandPath(ContainerPrefix, PathName());
       if (!ExpandedString.empty()) {
         // Adjust the path if it ended up being relative
         FEXCore::Config::EraseSet(FEXCore::Config::CONFIG_ROOTFS, ExpandedString);
@@ -309,7 +480,8 @@ namespace FEXCore::Config {
       else if (!PathName().empty()) {
         // If the filesystem doesn't exist then let's see if it exists in the fex-emu folder
         std::string NamedRootFS = GetDataDirectory() + "RootFS/" + PathName();
-        if (std::filesystem::exists(NamedRootFS)) {
+        std::error_code ec{};
+        if (std::filesystem::exists(NamedRootFS, ec)) {
           FEXCore::Config::EraseSet(FEXCore::Config::CONFIG_ROOTFS, NamedRootFS);
         }
       }
@@ -324,7 +496,19 @@ namespace FEXCore::Config {
     }
     if (FEXCore::Config::Exists(FEXCore::Config::CONFIG_THUNKCONFIG)) {
       FEX_CONFIG_OPT(PathName, THUNKCONFIG);
-      ExpandPathIfExists(FEXCore::Config::CONFIG_THUNKCONFIG, PathName());
+      auto ExpandedString = ExpandPath(ContainerPrefix, PathName());
+      if (!ExpandedString.empty()) {
+        // Adjust the path if it ended up being relative
+        FEXCore::Config::EraseSet(FEXCore::Config::CONFIG_THUNKCONFIG, ExpandedString);
+      }
+      else if (!PathName().empty()) {
+        // If the filesystem doesn't exist then let's see if it exists in the fex-emu folder
+        std::string NamedConfig = GetDataDirectory() + "ThunkConfigs/" + PathName();
+        std::error_code ec{};
+        if (std::filesystem::exists(NamedConfig, ec)) {
+          FEXCore::Config::EraseSet(FEXCore::Config::CONFIG_THUNKCONFIG, NamedConfig);
+        }
+      }
     }
     if (FEXCore::Config::Exists(FEXCore::Config::CONFIG_OUTPUTLOG)) {
       FEX_CONFIG_OPT(PathName, OUTPUTLOG);
@@ -402,6 +586,17 @@ namespace FEXCore::Config {
     }
   }
 
+  template<>
+  std::string Value<std::string>::GetIfExists(FEXCore::Config::ConfigOption Option, std::string_view Default) {
+    auto Value = FEXCore::Config::Get(Option);
+    if (Value) {
+      return **Value;
+    }
+    else {
+      return std::string(Default);
+    }
+  }
+
   template bool     Value<bool>::GetIfExists(FEXCore::Config::ConfigOption Option, bool Default);
   template int8_t   Value<int8_t>::GetIfExists(FEXCore::Config::ConfigOption Option, int8_t Default);
   template uint8_t  Value<uint8_t>::GetIfExists(FEXCore::Config::ConfigOption Option, uint8_t Default);
@@ -427,5 +622,143 @@ namespace FEXCore::Config {
     }
   }
   template void Value<std::string>::GetListIfExists(FEXCore::Config::ConfigOption Option, std::list<std::string> *List);
+
+  // Application loaders
+  class MainLoader final : public FEXCore::Config::OptionMapper {
+  public:
+    explicit MainLoader();
+    explicit MainLoader(std::string ConfigFile);
+    void Load() override;
+
+  private:
+    std::string Config;
+  };
+
+  class AppLoader final : public FEXCore::Config::OptionMapper {
+  public:
+    explicit AppLoader(const std::string& Filename, bool Global);
+    void Load();
+
+  private:
+    std::string Config;
+  };
+
+  class EnvLoader final : public FEXCore::Config::Layer {
+  public:
+    explicit EnvLoader(char *const _envp[]);
+    void Load() override;
+
+  private:
+    char *const *envp;
+  };
+
+  static const std::map<std::string, FEXCore::Config::ConfigOption, std::less<>> ConfigLookup = {{
+#define OPT_BASE(type, group, enum, json, default) {#json, FEXCore::Config::ConfigOption::CONFIG_##enum},
+#include <FEXCore/Config/ConfigValues.inl>
+  }};
+  static const std::vector<std::pair<const char*, FEXCore::Config::ConfigOption>> EnvConfigLookup = {{
+#define OPT_BASE(type, group, enum, json, default) {"FEX_" #enum, FEXCore::Config::ConfigOption::CONFIG_##enum},
+#include <FEXCore/Config/ConfigValues.inl>
+  }};
+
+  OptionMapper::OptionMapper(FEXCore::Config::LayerType Layer)
+    : FEXCore::Config::Layer(Layer) {
+  }
+
+  void OptionMapper::MapNameToOption(const char *ConfigName, const char *ConfigString) {
+    auto it = ConfigLookup.find(ConfigName);
+    if (it != ConfigLookup.end()) {
+      Set(it->second, ConfigString);
+    }
+  }
+
+  MainLoader::MainLoader()
+    : FEXCore::Config::OptionMapper(FEXCore::Config::LayerType::LAYER_MAIN)
+    , Config{FEXCore::Config::GetConfigFileLocation()} {
+  }
+
+  MainLoader::MainLoader(std::string ConfigFile)
+    : FEXCore::Config::OptionMapper(FEXCore::Config::LayerType::LAYER_MAIN)
+    , Config{std::move(ConfigFile)} {
+  }
+
+  void MainLoader::Load() {
+    JSON::LoadJSonConfig(Config, [this](const char *Name, const char *ConfigString) {
+      MapNameToOption(Name, ConfigString);
+    });
+  }
+
+  AppLoader::AppLoader(const std::string& Filename, bool Global)
+    : FEXCore::Config::OptionMapper(Global ? FEXCore::Config::LayerType::LAYER_GLOBAL_APP : FEXCore::Config::LayerType::LAYER_LOCAL_APP) {
+    Config = FEXCore::Config::GetApplicationConfig(Filename, Global);
+
+    // Immediately load so we can reload the meta layer
+    Load();
+  }
+
+  void AppLoader::Load() {
+    JSON::LoadJSonConfig(Config, [this](const char *Name, const char *ConfigString) {
+      MapNameToOption(Name, ConfigString);
+    });
+  }
+
+  EnvLoader::EnvLoader(char *const _envp[])
+    : FEXCore::Config::Layer(FEXCore::Config::LayerType::LAYER_ENVIRONMENT)
+    , envp {_envp} {
+  }
+
+  void EnvLoader::Load() {
+    std::unordered_map<std::string_view, std::string_view> EnvMap;
+
+    for(const char *const *pvar=envp; pvar && *pvar; pvar++) {
+      std::string_view Var(*pvar);
+      size_t pos = Var.rfind('=');
+      if (std::string::npos == pos)
+        continue;
+
+      std::string_view Ident = Var.substr(0,pos);
+      std::string_view Value = Var.substr(pos+1);
+      EnvMap[Ident]=Value;
+    }
+
+    std::function GetVar = [=](const std::string_view id)  -> std::optional<std::string_view> {
+      if (EnvMap.find(id) != EnvMap.end())
+        return EnvMap.at(id);
+
+      // If envp[] was empty, search using std::getenv()
+      const char* vs = std::getenv(id.data());
+      if (vs) {
+        return vs;
+      }
+      else {
+        return std::nullopt;
+      }
+    };
+
+    std::optional<std::string_view> Value;
+
+    for (auto &it : EnvConfigLookup) {
+      if ((Value = GetVar(it.first)).has_value()) {
+        Set(it.second, std::string(*Value));
+      }
+    }
+  }
+
+  std::unique_ptr<FEXCore::Config::Layer> CreateMainLayer(std::string const *File) {
+    if (File) {
+      return std::make_unique<FEXCore::Config::MainLoader>(*File);
+    }
+    else {
+      return std::make_unique<FEXCore::Config::MainLoader>();
+    }
+  }
+
+  std::unique_ptr<FEXCore::Config::Layer> CreateAppLayer(const std::string& Filename, bool Global) {
+    return std::make_unique<FEXCore::Config::AppLoader>(Filename, Global);
+  }
+
+  std::unique_ptr<FEXCore::Config::Layer> CreateEnvironmentLayer(char *const _envp[]) {
+    return std::make_unique<FEXCore::Config::EnvLoader>(_envp);
+  }
 }
 

External/FEXCore/Source/Interface/Config/Config.json.in

@@ -58,7 +58,7 @@
       },
       "ThunkHostLibs": {
         "Type": "str",
-        "Default": "",
+        "Default": "@CMAKE_INSTALL_PREFIX@/lib/fex-emu/HostThunks/",
         "ShortArg": "t",
         "Desc": [
           "Folder to find the host-side thunking libraries."
@@ -66,7 +66,7 @@
       },
       "ThunkGuestLibs": {
         "Type": "str",
-        "Default": "",
+        "Default": "@CMAKE_INSTALL_PREFIX@/share/fex-emu/GuestThunks/",
         "ShortArg": "j",
         "Desc": [
           "Folder to find the guest-side thunking libraries."
@@ -77,7 +77,14 @@
         "Default": "",
         "ShortArg": "k",
         "Desc": [
-          "A json file specifying where to overlay the thunks."
+          "A json file specifying where to overlay the thunks.",
+          "This can be a filesystem path",
+          "\teg: ~/MyThunkConfig.json",
+          "Or this can be a named of a Thunk config file",
+          "If the named config file exists in the FEX data folder folder the it will use that one",
+          "\teg: $HOME/.fex-emu/ThunkConfigs/<ThunkConfig name>",
+          "Or if you have XDG_DATA_HOME the config will search in that directory",
+          "\teg: $XDG_DATA_HOME/.fex-emu/ThunkConfigs/<ThunkConfig name>"
         ]
       },
       "Env": {
@@ -87,6 +94,16 @@
         "Desc": [
           "Adds an environment variable to the emulated environment."
         ]
+      },
+      "HostEnv": {
+        "Type": "strarray",
+        "Default": "",
+        "ShortArg": "H",
+        "Desc": [
+          "Adds an environment variable to the host environment.",
+          "This can be useful for setting environment variables that thunks can pick up.",
+          "Typically isn't necessary since the guest libc isn't thunked. But is possible."
+        ]
       }
     },
     "Debug": {
@@ -130,6 +147,13 @@
           "Disables optimizations passes for debugging."
         ]
       },
+      "SRA": {
+        "Type": "bool",
+        "Default": "true",
+        "Desc": [
+          "Set to false to disable Static Register Allocation"
+        ]
+      },
       "Force32BitAllocator": {
         "Type": "bool",
         "Default": "false",
@@ -207,6 +231,14 @@
           "Makes TSO operations even more strict.",
           "Forces vector loadstores to also become atomic."
         ]
+      },
+      "StallProcess": {
+        "Type": "bool",
+        "Default": "false",
+        "Desc": [
+          "Forces a process to stall out on initialization",
+          "Useful for a process that keeps restarting and doesn't work"
+        ]
       }
     },
     "Misc": {

External/FEXCore/Source/Interface/Context/Context.cpp

@@ -2,10 +2,20 @@
 #include "Interface/Context/Context.h"
 #include "Interface/Core/Core.h"
 #include "Interface/Core/OpcodeDispatcher.h"
+#include "Interface/Core/X86Tables/X86Tables.h"
 
-#include <FEXCore/Config/Config.h>
 #include <FEXCore/Core/CoreState.h>
-#include <FEXCore/Debug/X86Tables.h>
+#include <FEXCore/Core/Context.h>
+#include <FEXCore/Core/CPUID.h>
+#include <FEXCore/Core/SignalDelegator.h>
+#include "FEXCore/Debug/InternalThreadState.h"
+
+#include <string.h>
+#include <utility>
+
+namespace FEXCore::HLE {
+  class SyscallVisitor;
+}
 
 namespace FEXCore::Context {
   void InitializeStaticTables(OperatingMode Mode) {
@@ -33,7 +43,7 @@ namespace FEXCore::Context {
     delete CTX;
   }
 
-  bool InitCore(FEXCore::Context::Context *CTX, FEXCore::CodeLoader *Loader) {
+  FEXCore::Core::InternalThreadState* InitCore(FEXCore::Context::Context *CTX, FEXCore::CodeLoader *Loader) {
     return CTX->InitCore(Loader);
   }
 
@@ -100,8 +110,8 @@ namespace FEXCore::Context {
   void RegisterExternalSyscallVisitor(FEXCore::Context::Context *CTX, [[maybe_unused]] uint64_t Syscall, [[maybe_unused]] FEXCore::HLE::SyscallVisitor *Visitor) {
   }
 
-  void HandleCallback(FEXCore::Context::Context *CTX, uint64_t RIP) {
-    CTX->HandleCallback(RIP);
+  void HandleCallback(FEXCore::Context::Context *CTX, FEXCore::Core::InternalThreadState *Thread, uint64_t RIP) {
+    CTX->HandleCallback(Thread, RIP);
   }
 
   void RegisterHostSignalHandler(FEXCore::Context::Context *CTX, int Signal, HostSignalDelegatorFunction Func, bool Required) {
@@ -116,6 +126,10 @@ namespace FEXCore::Context {
     return CTX->CreateThread(NewThreadState, ParentTID);
   }
 
+  void ExecutionThread(FEXCore::Context::Context *CTX, FEXCore::Core::InternalThreadState *Thread) {
+    return CTX->ExecutionThread(Thread);
+  }
+
   void InitializeThread(FEXCore::Context::Context *CTX, FEXCore::Core::InternalThreadState *Thread) {
     return CTX->InitializeThread(Thread);
   }

External/FEXCore/Source/Interface/Context/Context.h

@@ -1,48 +1,51 @@
 #pragma once
 
-#include "Common/JitSymbols.h"
 #include "Interface/Core/CPUID.h"
-#include "Interface/Core/Frontend.h"
 #include "Interface/Core/HostFeatures.h"
-#include "Interface/Core/InternalThreadState.h"
 #include "Interface/Core/X86HelperGen.h"
-#include "Interface/IR/PassManager.h"
-#include "Interface/IR/Passes/RegisterAllocationPass.h"
 #include <FEXCore/Config/Config.h>
-#include <FEXCore/Core/CPUBackend.h>
+#include <FEXCore/Core/Context.h>
+#include <FEXCore/Core/CoreState.h>
+#include <FEXCore/Core/SignalDelegator.h>
+#include <FEXCore/Debug/InternalThreadState.h>
 #include <FEXCore/Utils/CompilerDefs.h>
 #include <FEXCore/Utils/Event.h>
 #include <stdint.h>
 
+#ifdef ENABLE_JITSYMBOLS
+#include <Common/JITSymbols.h>
+#endif
+
+#include <atomic>
+#include <condition_variable>
 #include <functional>
 #include <istream>
 #include <map>
 #include <memory>
 #include <mutex>
-#include <optional>
-#include <ostream>
-#include <set>
 #include <shared_mutex>
+#include <stddef.h>
+#include <string>
 #include <unordered_map>
 #include <queue>
+#include <vector>
 
 namespace FEXCore {
+class CodeLoader;
 class ThunkHandler;
-class BlockSamplingData;
 class GdbServer;
-class SiganlDelegator;
 
 namespace CPU {
   class Arm64JITCore;
   class X86JITCore;
 }
 namespace HLE {
+struct SyscallArguments;
 class SyscallHandler;
 }
 }
 
 namespace FEXCore::IR {
-  class RegisterAllocationPass;
   class RegisterAllocationData;
   class IRListView;
 namespace Validation {
@@ -121,7 +124,9 @@ namespace FEXCore::Context {
       FEX_CONFIG_OPT(MaxInstPerBlock, MAXINST);
       FEX_CONFIG_OPT(RootFSPath, ROOTFS);
       FEX_CONFIG_OPT(ThunkHostLibsPath, THUNKHOSTLIBS);
+      FEX_CONFIG_OPT(ThunkConfigFile, THUNKCONFIG);
       FEX_CONFIG_OPT(DumpIR, DUMPIR);
+      FEX_CONFIG_OPT(StaticRegisterAllocation, SRA);
     } Config;
 
     using IntCallbackReturn =  FEX_NAKED void(*)(FEXCore::Core::InternalThreadState *Thread, volatile void *Host_RSP);
@@ -183,7 +188,7 @@ namespace FEXCore::Context {
     Context();
     ~Context();
 
-    bool InitCore(FEXCore::CodeLoader *Loader);
+    FEXCore::Core::InternalThreadState* InitCore(FEXCore::CodeLoader *Loader);
     FEXCore::Context::ExitReason RunUntilExit();
     int GetProgramStatus() const;
     bool IsPaused() const { return !Running; }
@@ -199,7 +204,7 @@ namespace FEXCore::Context {
     bool GetGdbServerStatus() const { return DebugServer != nullptr; }
     void StartGdbServer();
     void StopGdbServer();
-    void HandleCallback(uint64_t RIP);
+    void HandleCallback(FEXCore::Core::InternalThreadState *Thread, uint64_t RIP);
     void RegisterHostSignalHandler(int Signal, HostSignalDelegatorFunction Func, bool Required);
     void RegisterFrontendHostSignalHandler(int Signal, HostSignalDelegatorFunction Func, bool Required);
 
@@ -240,23 +245,80 @@ namespace FEXCore::Context {
     };
     [[nodiscard]] CompileCodeResult CompileCode(FEXCore::Core::InternalThreadState *Thread, uint64_t GuestRIP);
     uintptr_t CompileBlock(FEXCore::Core::CpuStateFrame *Frame, uint64_t GuestRIP);
-    
+
     // same as CompileBlock, but aborts on failure
     void CompileBlockJit(FEXCore::Core::CpuStateFrame *Frame, uint64_t GuestRIP);
 
     bool LoadAOTIRCache(int streamfd);
     void FinalizeAOTIRCache();
     void WriteFilesWithCode(std::function<void(const std::string& fileid, const std::string& filename)> Writer);
-    
-    // Used for thread creation from syscalls
+    /**
+     * @brief Initializes the JIT compilers for the thread
+     *
+     * @param State The internal FEX thread state object
+     * @param CompileThread Is this for the compile service or not?
+     *
+     * InitializeCompiler is called inside of CreateThread, so you likely don't need this
+     * This is exposed because the CompileService needs to initialize compilers while copying data from
+     * the paired InternalThreadState that it is compiling code for
+     */
     void InitializeCompiler(FEXCore::Core::InternalThreadState* State, bool CompileThread);
+
+    // Used for thread creation from syscalls
+    /**
+     * @brief Used to create FEX thread objects in preparation for creating a true OS thread
+     *
+     * @param NewThreadState The initial thread state to setup for our state
+     * @param ParentTID The PID that was the parent thread that created this
+     *
+     * @return The InternalThreadState object that tracks all of the emulated thread's state
+     *
+     * Usecases:
+     *  OS thread Creation:
+     *    - Thread = CreateThread(NewState, PPID);
+     *    - InitializeThread(Thread);
+     *  OS fork (New thread created with a clone of thread state):
+     *    - clone{2, 3}
+     *    - Thread = CreateThread(CopyOfThreadState, PPID);
+     *    - ExecutionThread(Thread); // Starts executing without creating another host thread
+     *  Thunk callback executing guest code from native host thread
+     *    - Thread = CreateThread(NewState, PPID);
+     *    - InitializeThreadTLSData(Thread);
+     *    - HandleCallback(Thread, RIP);
+     */
     FEXCore::Core::InternalThreadState* CreateThread(FEXCore::Core::CPUState *NewThreadState, uint64_t ParentTID);
-    void InitializeThreadData(FEXCore::Core::InternalThreadState *Thread);
+
+    /**
+     * @brief Initializes the TLS data for a thread
+     *
+     * @param Thread The internal FEX thread state object
+     */
+    void InitializeThreadTLSData(FEXCore::Core::InternalThreadState *Thread);
+
+    /**
+     * @brief Initializes the OS thread object and prepares to start executing on that new OS thread
+     *
+     * @param Thread The internal FEX thread state object
+     *
+     * The OS thread will wait until RunThread is executed
+     */
     void InitializeThread(FEXCore::Core::InternalThreadState *Thread);
-    void CopyMemoryMapping(FEXCore::Core::InternalThreadState *ParentThread, FEXCore::Core::InternalThreadState *ChildThread);
+
+    /**
+     * @brief Starts the OS thread object to start executing guest code
+     *
+     * @param Thread The internal FEX thread state object
+     */
     void RunThread(FEXCore::Core::InternalThreadState *Thread);
 
+    /**
+     * @brief Destroys this FEX thread object and stops tracking it internally
+     *
+     * @param Thread The internal FEX thread state object
+     */
     void DestroyThread(FEXCore::Core::InternalThreadState *Thread);
+    void CopyMemoryMapping(FEXCore::Core::InternalThreadState *ParentThread, FEXCore::Core::InternalThreadState *ChildThread);
+
     void CleanupAfterFork(FEXCore::Core::InternalThreadState *ExceptForThread);
 
     std::vector<FEXCore::Core::InternalThreadState*>* GetThreads() { return &Threads; }
@@ -277,6 +339,15 @@ namespace FEXCore::Context {
     void ClearCodeCache(FEXCore::Core::InternalThreadState *Thread, bool AlsoClearIRCache);
 
   private:
+    /**
+     * @brief Does some final thread initialization
+     *
+     * @param Thread The internal FEX thread state object
+     *
+     * InitCore and CreateThread both call this to finish up thread object initialization
+     */
+    void InitializeThreadData(FEXCore::Core::InternalThreadState *Thread);
+
     void WaitForIdleWithTimeout();
 
     void NotifyPause();

External/FEXCore/Source/Interface/Core/ArchHelpers/Arm64.cpp

@@ -1,6 +1,8 @@
 #include "Interface/Core/ArchHelpers/Arm64.h"
+#include "Interface/Core/ArchHelpers/MContext.h"
 
 #include <FEXCore/Utils/LogManager.h>
+#include <FEXCore/Utils/Telemetry.h>
 
 #include <atomic>
 #include <stdint.h>
@@ -8,6 +10,30 @@
 #include <signal.h>
 
 namespace FEXCore::ArchHelpers::Arm64 {
+FEXCORE_TELEMETRY_STATIC_INIT(SplitLock, TYPE_HAS_SPLIT_LOCKS);
+FEXCORE_TELEMETRY_STATIC_INIT(SplitLock16B, TYPE_16BYTE_SPLIT);
+
+static __uint128_t LoadAcquire128(uint64_t Addr) {
+  __uint128_t Result{};
+  uint64_t Lower;
+  uint64_t Upper;
+  // This specifically avoids using std::atomic<__uint128_t>
+  // std::atomic helper does a ldaxp + stxp pair that crashes when the page is only mapped readable
+  __asm volatile(
+R"(
+  ldaxp %[ResultLower], %[ResultUpper], [%[Addr]];
+  clrex;
+)"
+  : [ResultLower] "=r" (Lower)
+  , [ResultUpper] "=r" (Upper)
+  : [Addr] "r" (Addr)
+  : "memory");
+  Result = Upper;
+  Result <<= 64;
+  Result |= Lower;
+  return Result;
+}
+
 static uint64_t LoadAcquire64(uint64_t Addr) {
   std::atomic<uint64_t> *Atom = reinterpret_cast<std::atomic<uint64_t>*>(Addr);
   return Atom->load(std::memory_order_acquire);
@@ -38,22 +64,12 @@ static bool StoreCAS8(uint8_t &Expected, uint8_t Val, uint64_t Addr) {
   return Atom->compare_exchange_strong(Expected, Val);
 }
 
-bool HandleCASPAL(void *_ucontext, void *_info, uint32_t Instr) {
+
+
+static bool RunCASPAL(void *_ucontext, void *_info, uint32_t Size, uint32_t DesiredReg1, uint32_t DesiredReg2, uint32_t ExpectedReg1, uint32_t ExpectedReg2, uint32_t AddressReg) {
   mcontext_t* mcontext = &reinterpret_cast<ucontext_t*>(_ucontext)->uc_mcontext;
-  siginfo_t* info = reinterpret_cast<siginfo_t*>(_info);
-
-  if (info->si_code != BUS_ADRALN) {
-    // This only handles alignment problems
-    return false;
-  }
-
-  uint32_t Size = (Instr >> 30) & 1;
-
-  uint32_t DesiredReg1 = Instr & 0b11111;
-  uint32_t DesiredReg2 = DesiredReg1 + 1;
-  uint32_t ExpectedReg1 = (Instr >> 16) & 0b11111;
-  uint32_t ExpectedReg2 = ExpectedReg1 + 1;
-  uint32_t AddressReg = (Instr >> 5) & 0b11111;
+  
+  //Bus_ADRALN check happens in HandleCASPAL and HandleCASPAL_ARMv8
 
   if (Size == 0) {
     // 32bit
@@ -72,8 +88,15 @@ bool HandleCASPAL(void *_ucontext, void *_info, uint32_t Instr) {
     // Both cross-cacheline and cross 16byte both need dual CAS loops that can tear
     // ARMv8.4 LSE2 solves all atomic issues except cross-cacheline
 
+    // Check for Split lock across a cacheline
+    if ((Addr & 63) > 56) {
+      FEXCORE_TELEMETRY_SET(SplitLock, 1);
+    }
+
     uint64_t AlignmentMask = 0b1111;
     if ((Addr & AlignmentMask) > 8) {
+      FEXCORE_TELEMETRY_SET(SplitLock16B, 1);
+
       uint64_t Alignment = Addr & 0b111;
       Addr &= ~0b111ULL;
       uint64_t AddrUpper = Addr + 8;
@@ -226,6 +249,256 @@ bool HandleCASPAL(void *_ucontext, void *_info, uint32_t Instr) {
   return false;
 }
 
+bool HandleCASPAL(void *_ucontext, void *_info, uint32_t Instr) {
+  siginfo_t* info = reinterpret_cast<siginfo_t*>(_info);
+
+  if (info->si_code != BUS_ADRALN) {
+    // This only handles alignment problems
+    return false;
+  }
+
+  uint32_t Size = (Instr >> 30) & 1;
+
+  uint32_t DesiredReg1 = Instr & 0b11111;
+  uint32_t DesiredReg2 = DesiredReg1 + 1;
+  uint32_t ExpectedReg1 = (Instr >> 16) & 0b11111;
+  uint32_t ExpectedReg2 = ExpectedReg1 + 1;
+  uint32_t AddressReg = (Instr >> 5) & 0b11111;
+  
+  return RunCASPAL(_ucontext, _info, Size, DesiredReg1, DesiredReg2, ExpectedReg1, ExpectedReg2, AddressReg);
+}
+
+uint64_t HandleCASPAL_ARMv8(void *_ucontext, void *_info, uint32_t Instr) {
+  mcontext_t* mcontext = &reinterpret_cast<ucontext_t*>(_ucontext)->uc_mcontext;
+  siginfo_t* info = reinterpret_cast<siginfo_t*>(_info);
+
+  if (info->si_code != BUS_ADRALN) {
+    // This only handles alignment problems
+    return 0;
+  }
+  // caspair
+  // [1] ldaxp(TMP2.W(), TMP3.W(), MemOperand(MemSrc)); <-- DataReg & AddrReg
+  // [2] cmp(TMP2.W(), Expected.first.W()); <-- ExpectedReg1
+  // [3] ccmp(TMP3.W(), Expected.second.W(), NoFlag, Condition::eq); <-- ExpectedREg2
+  // [4] b(&LoopNotExpected, Condition::ne);
+  // [5] stlxp(TMP2.W(), Desired.first.W(), Desired.second.W(), MemOperand(MemSrc)); <-- DesiredReg
+  // [6] cbnz(TMP2.W(), &LoopTop);
+  // [7] mov(Dst.first.W(), Expected.first.W());
+  // [8] mov(Dst.second.W(), Expected.second.W());
+  // [9] b(&LoopExpected);
+  // [10] mov(Dst.first.W(), TMP2.W());
+  // [11] mov(Dst.second.W(), TMP3.W());
+  // [12] clrex();
+
+  uint32_t *PC = (uint32_t*)ArchHelpers::Context::GetPc(_ucontext);
+  
+  uint32_t Size = (Instr >> 30) & 1;
+  uint32_t AddrReg = (Instr >> 5) & 0x1F;
+  uint32_t DataReg = Instr & 0x1F;
+  uint32_t DataReg2 = (Instr >> 10) & 0x1F;
+  
+  uint32_t ExpectedReg1{};
+  uint32_t ExpectedReg2{};
+  
+  uint32_t DesiredReg1{};
+  uint32_t DesiredReg2{};
+  
+  if(Size != 0) { //Only 32-bit pairs
+    return 0; 
+  }
+  
+  for(int i = 1; i < 10; i++) {
+    uint32_t NextInstr = PC[i];
+    if ((NextInstr & FEXCore::ArchHelpers::Arm64::ALU_OP_MASK) == FEXCore::ArchHelpers::Arm64::CMP_INST) {
+       ExpectedReg1 = GetRmReg(NextInstr);
+    } else if ((NextInstr & FEXCore::ArchHelpers::Arm64::CCMP_MASK) == FEXCore::ArchHelpers::Arm64::CCMP_INST) {
+       ExpectedReg2 = GetRmReg(NextInstr);
+    } else if ((NextInstr & FEXCore::ArchHelpers::Arm64::STLXP_MASK) == FEXCore::ArchHelpers::Arm64::STLXP_INST) {
+       DesiredReg1 = (NextInstr & 0x1F);
+       DesiredReg2 = (NextInstr >> 10) & 0x1F;
+    }
+  }
+  
+  //mov expected into the temp registers used by JIT
+  mcontext->regs[DataReg] = mcontext->regs[ExpectedReg1];
+  mcontext->regs[DataReg2] = mcontext->regs[ExpectedReg2];
+  
+  if(RunCASPAL(_ucontext, _info, Size, DesiredReg1, DesiredReg2, DataReg, DataReg2, AddrReg)) {
+    return 9 * sizeof(uint32_t); // skip to mov + clrex
+  } else {
+    return 0;
+  }
+}
+
+uint16_t DoLoad16(uint64_t Addr) {
+  uint64_t AlignmentMask = 0b1111;
+  if ((Addr & AlignmentMask) == 15) {
+    // Address crosses over 16byte or 64byte threshold
+    // Needs two loads
+    uint64_t AddrUpper = Addr + 1;
+    uint8_t ActualUpper{};
+    uint8_t ActualLower{};
+    // Careful ordering here
+    ActualUpper = LoadAcquire8(AddrUpper);
+    ActualLower = LoadAcquire8(Addr);
+
+    uint16_t Result = ActualUpper;
+    Result <<= 8;
+    Result |= ActualLower;
+    return Result;
+  }
+  else {
+    AlignmentMask = 0b111;
+    if ((Addr & AlignmentMask) == 7) {
+      // Crosses 8byte boundary
+      // Needs 128bit load
+      // Fits within a 16byte region
+      uint64_t Alignment = Addr & 0b1111;
+      Addr &= ~0b1111ULL;
+
+      __uint128_t TmpResult = LoadAcquire128(Addr);
+
+      // Zexts the result
+      uint16_t Result = TmpResult >> (Alignment * 8);
+      return Result;
+    }
+    else {
+      AlignmentMask = 0b11;
+      if ((Addr & AlignmentMask) == 3) {
+        // Crosses 4byte boundary
+        // Needs 64bit Load
+        uint64_t Alignment = Addr & AlignmentMask;
+        Addr &= ~AlignmentMask;
+
+        std::atomic<uint64_t> *Atomic = reinterpret_cast<std::atomic<uint64_t>*>(Addr);
+        uint64_t TmpResult = Atomic->load();
+
+        // Zexts the result
+        uint16_t Result = TmpResult >> (Alignment * 8);
+        return Result;
+      }
+      else {
+        // Fits within 4byte boundary
+        // Only needs 32bit Load
+        // Only alignment offset will be 1 here
+        uint64_t Alignment = Addr & AlignmentMask;
+        Addr &= ~AlignmentMask;
+
+        std::atomic<uint32_t> *Atomic = reinterpret_cast<std::atomic<uint32_t>*>(Addr);
+        uint32_t TmpResult = Atomic->load();
+
+        // Zexts the result
+        uint16_t Result = TmpResult >> (Alignment * 8);
+        return Result;
+      }
+    }
+  }
+}
+
+uint32_t DoLoad32(uint64_t Addr) {
+  uint64_t AlignmentMask = 0b1111;
+  if ((Addr & AlignmentMask) > 12) {
+    // Address crosses over 16byte threshold
+    // Needs dual 32bit load
+    uint64_t Alignment = Addr & 0b11;
+    Addr &= ~0b11ULL;
+
+    uint64_t AddrUpper = Addr + 4;
+
+    // Careful ordering here
+    uint32_t ActualUpper = LoadAcquire32(AddrUpper);
+    uint32_t ActualLower = LoadAcquire32(Addr);
+
+    uint64_t Result = ActualUpper;
+    Result <<= 32;
+    Result |= ActualLower;
+    return Result >> (Alignment * 8);
+  }
+  else {
+    AlignmentMask = 0b111;
+    if ((Addr & AlignmentMask) >= 5) {
+      // Crosses 8byte boundary
+      // Needs 128bit load
+      // Fits within a 16byte region
+      uint64_t Alignment = Addr & 0b1111;
+      Addr &= ~0b1111ULL;
+
+      __uint128_t TmpResult = LoadAcquire128(Addr);
+
+      return TmpResult >> (Alignment * 8);
+    }
+    else {
+      // Fits within 8byte boundary
+      // Only needs 64bit CAS
+      // Alignments can be [1,5)
+      uint64_t Alignment = Addr & AlignmentMask;
+      Addr &= ~AlignmentMask;
+
+      std::atomic<uint64_t> *Atomic = reinterpret_cast<std::atomic<uint64_t>*>(Addr);
+      uint64_t TmpResult = Atomic->load();
+
+      return TmpResult >> (Alignment * 8);
+    }
+  }
+}
+
+uint64_t DoLoad64(uint64_t Addr) {
+  uint64_t AlignmentMask = 0b1111;
+  if ((Addr & AlignmentMask) > 8) {
+    uint64_t Alignment = Addr & 0b111;
+    Addr &= ~0b111ULL;
+    uint64_t AddrUpper = Addr + 8;
+
+    // Crosses a 16byte boundary
+    // Needs two 8 byte loads
+    uint64_t ActualUpper{};
+    uint64_t ActualLower{};
+    // Careful ordering here
+    ActualUpper = LoadAcquire64(AddrUpper);
+    ActualLower = LoadAcquire64(Addr);
+
+    __uint128_t Result = ActualUpper;
+    Result <<= 64;
+    Result |= ActualLower;
+    return Result >> (Alignment * 8);
+  }
+  else {
+    // Fits within a 16byte region
+    uint64_t Alignment = Addr & AlignmentMask;
+    Addr &= ~AlignmentMask;
+    __uint128_t TmpResult = LoadAcquire128(Addr);
+    uint64_t Result = TmpResult >> (Alignment * 8);
+    return Result;
+  }
+}
+
+std::pair<uint64_t, uint64_t> DoLoad128(uint64_t Addr) {
+  // Any misalignment here means we cross a 16byte boundary
+  // So we need two 128bit loads
+  uint64_t Alignment = Addr & 0b1111;
+  Addr &= ~0b1111ULL;
+  uint64_t AddrUpper = Addr + 16;
+
+  union AlignedData {
+    struct {
+      __uint128_t Lower;
+      __uint128_t Upper;
+    } Large;
+    struct {
+      uint8_t Data[32];
+    } Bytes;
+  };
+
+  AlignedData *Data = reinterpret_cast<AlignedData*>(alloca(sizeof(AlignedData)));
+  Data->Large.Upper = LoadAcquire128(AddrUpper);
+  Data->Large.Lower = LoadAcquire128(Addr);
+
+  uint64_t ResultLower{}, ResultUpper{};
+  memcpy(&ResultLower, &Data->Bytes.Data[Alignment], sizeof(uint64_t));
+  memcpy(&ResultUpper, &Data->Bytes.Data[Alignment + sizeof(uint64_t)], sizeof(uint64_t));
+  return {ResultLower, ResultUpper};
+}
+
 template <typename T>
 using CASExpectedFn = T (*)(T Src, T Expected);
 template <typename T>
@@ -239,9 +512,16 @@ uint16_t DoCAS16(
   uint64_t Addr,
   CASExpectedFn<uint16_t> ExpectedFunction,
   CASDesiredFn<uint16_t> DesiredFunction) {
+
+  if ((Addr & 63) == 63) {
+    FEXCORE_TELEMETRY_SET(SplitLock, 1);
+  }
+
   // 16 bit
   uint64_t AlignmentMask = 0b1111;
   if ((Addr & AlignmentMask) == 15) {
+    FEXCORE_TELEMETRY_SET(SplitLock16B, 1);
+
     // Address crosses over 16byte or 64byte threshold
     // Need a dual 8bit CAS loop
     uint64_t AddrUpper = Addr + 1;
@@ -515,9 +795,16 @@ uint32_t DoCAS32(
   uint64_t Addr,
   CASExpectedFn<uint32_t> ExpectedFunction,
   CASDesiredFn<uint32_t> DesiredFunction) {
+
+  if ((Addr & 63) > 60) {
+    FEXCORE_TELEMETRY_SET(SplitLock, 1);
+  }
+
   // 32 bit
   uint64_t AlignmentMask = 0b1111;
   if ((Addr & AlignmentMask) > 12) {
+    FEXCORE_TELEMETRY_SET(SplitLock16B, 1);
+
     // Address crosses over 16byte threshold
     // Needs dual 4 byte CAS loop
     uint64_t Alignment = Addr & 0b11;
@@ -745,9 +1032,16 @@ uint64_t DoCAS64(
   uint64_t Addr,
   CASExpectedFn<uint64_t> ExpectedFunction,
   CASDesiredFn<uint64_t> DesiredFunction) {
+
+  if ((Addr & 63) > 56) {
+    FEXCORE_TELEMETRY_SET(SplitLock, 1);
+  }
+
   // 64bit
   uint64_t AlignmentMask = 0b1111;
   if ((Addr & AlignmentMask) > 8) {
+    FEXCORE_TELEMETRY_SET(SplitLock16B, 1);
+
     uint64_t Alignment = Addr & 0b111;
     Addr &= ~0b111ULL;
     uint64_t AddrUpper = Addr + 8;
@@ -900,21 +1194,10 @@ uint64_t DoCAS64(
   }
 }
 
-bool HandleCASAL(void *_ucontext, void *_info, uint32_t Instr) {
+static bool RunCASAL(void *_ucontext, void *_info, uint32_t Size, uint32_t DesiredReg, uint32_t ExpectedReg, uint32_t AddressReg) {
   mcontext_t* mcontext = &reinterpret_cast<ucontext_t*>(_ucontext)->uc_mcontext;
   siginfo_t* info = reinterpret_cast<siginfo_t*>(_info);
 
-  if (info->si_code != BUS_ADRALN) {
-    // This only handles alignment problems
-    return false;
-  }
-
-  uint32_t Size = 1 << (Instr >> 30);
-
-  uint32_t DesiredReg = Instr & 0b11111;
-  uint32_t ExpectedReg = (Instr >> 16) & 0b11111;
-  uint32_t AddressReg = (Instr >> 5) & 0b11111;
-
   uint64_t Addr = mcontext->regs[AddressReg];
 
   // Cross-cacheline CAS doesn't work on ARM
@@ -994,6 +1277,23 @@ bool HandleCASAL(void *_ucontext, void *_info, uint32_t Instr) {
   return false;
 }
 
+bool HandleCASAL(void *_ucontext, void *_info, uint32_t Instr) {
+  mcontext_t* mcontext = &reinterpret_cast<ucontext_t*>(_ucontext)->uc_mcontext;
+  siginfo_t* info = reinterpret_cast<siginfo_t*>(_info);
+
+  if (info->si_code != BUS_ADRALN) {
+    // This only handles alignment problems
+    return false;
+  }
+
+  uint32_t Size = 1 << (Instr >> 30);
+
+  uint32_t DesiredReg = Instr & 0b11111;
+  uint32_t ExpectedReg = (Instr >> 16) & 0b11111;
+  uint32_t AddressReg = (Instr >> 5) & 0b11111;
+  return RunCASAL(_ucontext, _info, Size, DesiredReg, ExpectedReg, AddressReg);
+}
+
 bool HandleAtomicMemOp(void *_ucontext, void *_info, uint32_t Instr) {
   mcontext_t* mcontext = &reinterpret_cast<ucontext_t*>(_ucontext)->uc_mcontext;
   siginfo_t* info = reinterpret_cast<siginfo_t*>(_info);
@@ -1202,4 +1502,549 @@ bool HandleAtomicMemOp(void *_ucontext, void *_info, uint32_t Instr) {
   return false;
 }
 
+bool HandleAtomicLoad(void *_ucontext, void *_info, uint32_t Instr) {
+  mcontext_t* mcontext = &reinterpret_cast<ucontext_t*>(_ucontext)->uc_mcontext;
+  siginfo_t* info = reinterpret_cast<siginfo_t*>(_info);
+
+  if (info->si_code != BUS_ADRALN) {
+    // This only handles alignment problems
+    return false;
+  }
+  uint32_t Size = 1 << (Instr >> 30);
+
+  uint32_t ResultReg = Instr & 0b11111;
+  uint32_t AddressReg = (Instr >> 5) & 0b11111;
+
+  uint64_t Addr = mcontext->regs[AddressReg];
+
+  if (Size == 2) {
+    auto Res = DoLoad16(Addr);
+    // We set the result register if it isn't a zero register
+    if (ResultReg != 31) {
+      mcontext->regs[ResultReg] = Res;
+    }
+    return true;
+  }
+  else if (Size == 4) {
+    auto Res = DoLoad32(Addr);
+    // We set the result register if it isn't a zero register
+    if (ResultReg != 31) {
+      mcontext->regs[ResultReg] = Res;
+    }
+    return true;
+  }
+  else if (Size == 8) {
+    auto Res = DoLoad64(Addr);
+    // We set the result register if it isn't a zero register
+    if (ResultReg != 31) {
+      mcontext->regs[ResultReg] = Res;
+    }
+    return true;
+  }
+
+  return false;
+}
+
+bool HandleAtomicStore(void *_ucontext, void *_info, uint32_t Instr) {
+  mcontext_t* mcontext = &reinterpret_cast<ucontext_t*>(_ucontext)->uc_mcontext;
+  siginfo_t* info = reinterpret_cast<siginfo_t*>(_info);
+
+  if (info->si_code != BUS_ADRALN) {
+    // This only handles alignment problems
+    return false;
+  }
+  uint32_t Size = 1 << (Instr >> 30);
+
+  uint32_t DataReg = Instr & 0x1F;
+  uint32_t AddressReg = (Instr >> 5) & 0b11111;
+
+  uint64_t Addr = mcontext->regs[AddressReg];
+
+  constexpr bool DoRetry = false;
+  if (Size == 2) {
+    DoCAS16<DoRetry>(
+      mcontext->regs[DataReg],
+      0, // Unused
+      Addr,
+      [](uint16_t SrcVal, uint16_t) -> uint16_t {
+        // Expected is just src
+        return SrcVal;
+      },
+      [](uint16_t, uint16_t Desired) -> uint16_t {
+        // Desired is just Desired
+        return Desired;
+      });
+    return true;
+  }
+  else if (Size == 4) {
+    DoCAS32<DoRetry>(
+      mcontext->regs[DataReg],
+      0, // Unused
+      Addr,
+      [](uint32_t SrcVal, uint32_t) -> uint32_t {
+        // Expected is just src
+        return SrcVal;
+      },
+      [](uint32_t, uint32_t Desired) -> uint32_t {
+        // Desired is just Desired
+        return Desired;
+      });
+    return true;
+  }
+  else if (Size == 8) {
+    DoCAS64<DoRetry>(
+      mcontext->regs[DataReg],
+      0, // Unused
+      Addr,
+      [](uint64_t SrcVal, uint64_t) -> uint64_t {
+        // Expected is just src
+        return SrcVal;
+      },
+      [](uint64_t, uint64_t Desired) -> uint64_t {
+        // Desired is just Desired
+        return Desired;
+      });
+    return true;
+  }
+
+  return false;
+}
+
+bool HandleAtomicLoad128(void *_ucontext, void *_info, uint32_t Instr) {
+  mcontext_t* mcontext = &reinterpret_cast<ucontext_t*>(_ucontext)->uc_mcontext;
+  siginfo_t* info = reinterpret_cast<siginfo_t*>(_info);
+
+  if (info->si_code != BUS_ADRALN) {
+    // This only handles alignment problems
+    return false;
+  }
+  uint32_t ResultReg = Instr & 0b11111;
+  uint32_t ResultReg2 = (Instr >> 10) & 0x1F;
+  uint32_t AddressReg = (Instr >> 5) & 0b11111;
+
+  uint64_t Addr = mcontext->regs[AddressReg];
+
+  auto Res = DoLoad128(Addr);
+  // We set the result register if it isn't a zero register
+  if (ResultReg != 31) {
+    mcontext->regs[ResultReg] = std::get<0>(Res);
+  }
+  if (ResultReg2 != 31) {
+    mcontext->regs[ResultReg2] = std::get<1>(Res);
+  }
+
+  return true;
+}
+
+static uint64_t HandleCAS_NoAtomics(void *_ucontext, void *_info)
+{
+  mcontext_t* mcontext = &reinterpret_cast<ucontext_t*>(_ucontext)->uc_mcontext;
+  siginfo_t* info = reinterpret_cast<siginfo_t*>(_info);
+  
+  // ARMv8.0 CAS
+  // [1] ldaxrb(TMP2.W(), MemOperand(MemSrc))
+  // [2] cmp (TMP2.W(), Expected.W())
+  // [3] b
+  // [4] stlxrb(TMP3.W(), Desired.W(), MemOperand(MemSrc)
+  // [5] cbnz
+  // [6] mov
+  // [7] b
+  // [8] mov (.., TMP2.W());
+  // [9] clrex
+  
+  uint32_t *PC = (uint32_t*)ArchHelpers::Context::GetPc(_ucontext);
+  uint32_t Instr = PC[0];
+  uint32_t Size = 1 << (Instr >> 30);
+  uint32_t AddressReg = GetRnReg(Instr);
+  uint32_t ResultReg = GetRdReg(Instr); //TMP2 
+  uint32_t DesiredReg = 0;
+  uint32_t ExpectedReg = 0;
+  for (size_t i = 1; i < 6; ++i) {
+     uint32_t NextInstr = PC[i];
+     if ((NextInstr & FEXCore::ArchHelpers::Arm64::STLXR_MASK) == FEXCore::ArchHelpers::Arm64::STLXR_INST) {
+       #if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
+       // Just double check that the memory destination matches
+       uint32_t StoreAddressReg = GetRnReg(NextInstr);
+       LOGMAN_THROW_A(StoreAddressReg == AddressReg, "StoreExclusive memory register didn't match the store exclusive register");
+       #endif
+       DesiredReg = GetRdReg(NextInstr);
+     }
+     else if ((NextInstr & FEXCore::ArchHelpers::Arm64::ALU_OP_MASK) == FEXCore::ArchHelpers::Arm64::CMP_INST) {
+       ExpectedReg = GetRmReg(NextInstr);
+     }
+  }
+  //set up CASAL by doing mov(TMP2, Expected)
+  mcontext->regs[ResultReg] = mcontext->regs[ExpectedReg];
+  
+  if(RunCASAL(_ucontext, _info, Size, DesiredReg, ResultReg, AddressReg)) {
+    return 7 * sizeof(uint32_t); //jump to mov to allocated register
+  } else {
+    return 0;
+  }
+}
+
+uint64_t HandleAtomicLoadstoreExclusive(void *_ucontext, void *_info) {
+  mcontext_t* mcontext = &reinterpret_cast<ucontext_t*>(_ucontext)->uc_mcontext;
+  siginfo_t* info = reinterpret_cast<siginfo_t*>(_info);
+
+  if (info->si_code != BUS_ADRALN) {
+    // This only handles alignment problems
+    return 0;
+  }
+
+  uint32_t *PC = (uint32_t*)ArchHelpers::Context::GetPc(_ucontext);
+  uint32_t Instr = PC[0];
+
+  // Atomic Add
+  // [1] ldaxrb(TMP2.W(), MemOperand(MemSrc));
+  // [2] add(TMP2.W(), TMP2.W(), GetReg<RA_32>(Op->Header.Args[1].ID()));
+  // [3] stlxrb(TMP2.W(), TMP2.W(), MemOperand(MemSrc));
+  // [4] cbnz(TMP2.W(), &LoopTop);
+  //
+  // Atomic Fetch Add
+  // [1] ldaxrb(TMP2.W(), MemOperand(MemSrc));
+  // [2] add(TMP3.W(), TMP2.W(), GetReg<RA_32>(Op->Header.Args[1].ID()));
+  // [3] stlxrb(TMP4.W(), TMP3.W(), MemOperand(MemSrc));
+  // [4] cbnz(TMP4.W(), &LoopTop);
+  // [5] mov(GetReg<RA_32>(Node), TMP2.W());
+  //
+  // Atomic Swap
+  //
+  // [1] ldaxrb(TMP2.W(), MemOperand(MemSrc));
+  // [2] stlxrb(TMP4.W(), GetReg<RA_32>(Op->Header.Args[1].ID()), MemOperand(MemSrc));
+  // [3] cbnz(TMP4.W(), &LoopTop);
+  // [4] uxtb(GetReg<RA_64>(Node), TMP2.W());
+  //
+  // ASSUMPTIONS:
+  // - Both cases:
+  //   - The [2]ALU op: (Non NEG case)
+  //     - First source is from [1]ldaxr
+  //     - Second source is incoming value
+  //   - The [2]ALU op: (NEG case)
+  //     - First source is zero register
+  //     - The second source is the from [1]ldaxr
+  //   - No ALU op: (SWAP case)
+  //     - No DataSourceRegister
+  //
+  // - In Atomic case (non-fetch)
+  //   - The [3]stlxr instruction status + memory register are the SAME register
+  //
+  // - In Atomic FETCH case
+  //   - The [3]stlxr instruction's status + memory register are never the same register
+  //   - The [5]mov instruction source is always the destination register from [1] ldaxr*
+  uint32_t ResultReg = GetRdReg(Instr);
+  uint32_t AddressReg = GetRnReg(Instr);
+  uint64_t Addr = mcontext->regs[AddressReg];
+
+  size_t NumInstructionsToSkip = 0;
+
+  // Are we an Atomic op or AtomicFetch?
+  bool AtomicFetch = false;
+
+  // This is the register that is the incoming source to the ALU operation
+  // <DataResultReg> = <Load Exclusive Value> <Op> <DataSourceReg>
+  // NEG case is special
+  // <DataResultReg> = Zero <Sub> <Load Exclusive Value>
+  // DataSourceRegister must always be the Rm register
+  uint32_t DataSourceReg {};
+  ExclusiveAtomicPairType AtomicOp {ExclusiveAtomicPairType::TYPE_SWAP};
+
+  // Scan forward at most five instructions to find our instructions
+  for (size_t i = 1; i < 6; ++i) {
+    uint32_t NextInstr = PC[i];
+    if ((NextInstr & FEXCore::ArchHelpers::Arm64::ALU_OP_MASK) == FEXCore::ArchHelpers::Arm64::ADD_INST) {
+      AtomicOp = ExclusiveAtomicPairType::TYPE_ADD;
+      DataSourceReg = GetRmReg(NextInstr);
+    }
+    else if ((NextInstr & FEXCore::ArchHelpers::Arm64::ALU_OP_MASK) == FEXCore::ArchHelpers::Arm64::SUB_INST) {
+      uint32_t RnReg = GetRnReg(NextInstr);
+      if (RnReg == REGISTER_MASK) {
+        // Zero reg means neg
+        AtomicOp = ExclusiveAtomicPairType::TYPE_NEG;
+      }
+      else {
+        AtomicOp = ExclusiveAtomicPairType::TYPE_SUB;
+      }
+      DataSourceReg = GetRmReg(NextInstr);
+    }
+    else if ((NextInstr & FEXCore::ArchHelpers::Arm64::ALU_OP_MASK) == FEXCore::ArchHelpers::Arm64::CMP_INST) {
+    	return HandleCAS_NoAtomics(_ucontext, _info); //ARMv8.0 CAS
+    }
+    else if ((NextInstr & FEXCore::ArchHelpers::Arm64::ALU_OP_MASK) == FEXCore::ArchHelpers::Arm64::AND_INST) {
+      AtomicOp = ExclusiveAtomicPairType::TYPE_AND;
+      DataSourceReg = GetRmReg(NextInstr);
+    }
+    else if ((NextInstr & FEXCore::ArchHelpers::Arm64::ALU_OP_MASK) == FEXCore::ArchHelpers::Arm64::OR_INST) {
+      AtomicOp = ExclusiveAtomicPairType::TYPE_OR;
+      DataSourceReg = GetRmReg(NextInstr);
+    }
+    else if ((NextInstr & FEXCore::ArchHelpers::Arm64::ALU_OP_MASK) == FEXCore::ArchHelpers::Arm64::EOR_INST) {
+      AtomicOp = ExclusiveAtomicPairType::TYPE_EOR;
+      DataSourceReg = GetRmReg(NextInstr);
+    }
+    else if ((NextInstr & FEXCore::ArchHelpers::Arm64::STLXR_MASK) == FEXCore::ArchHelpers::Arm64::STLXR_INST) {
+#if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
+      // Just double check that the memory destination matches
+      uint32_t StoreAddressReg = GetRnReg(NextInstr);
+      LOGMAN_THROW_A(StoreAddressReg == AddressReg, "StoreExclusive memory register didn't match the store exclusive register");
+#endif
+      uint32_t StatusReg = GetRmReg(NextInstr);
+      uint32_t StoreResultReg = GetRdReg(NextInstr);
+      // We are an atomic fetch instruction if the data register isn't the status register
+      AtomicFetch = !(StatusReg == StoreResultReg);
+      if (AtomicOp == ExclusiveAtomicPairType::TYPE_SWAP) {
+        // In the case of swap we don't have an ALU op inbetween
+        // Source is directly in STLXR
+        DataSourceReg = StoreResultReg;
+      }
+    }
+    else if ((NextInstr & FEXCore::ArchHelpers::Arm64::CBNZ_MASK) == FEXCore::ArchHelpers::Arm64::CBNZ_INST) {
+      // Found the CBNZ, we want to skip to just after this instruction when done
+      NumInstructionsToSkip = i + 1;
+      // This is the last instruction we care about. Leave now
+      break;
+    }
+    else {
+      LogMan::Msg::A("Unknown instruction 0x%08x", NextInstr);
+    }
+  }
+
+  uint32_t Size = 1 << (Instr >> 30);
+
+  constexpr bool DoRetry = true;
+  if (Size == 2) {
+    using AtomicType = uint16_t;
+    auto NOPExpected = [](AtomicType SrcVal, AtomicType) -> AtomicType {
+      return SrcVal;
+    };
+
+    auto ADDDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal + Desired;
+    };
+
+    auto SUBDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal - Desired;
+    };
+
+    auto ANDDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal & Desired;
+    };
+
+    auto ORDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal | Desired;
+    };
+
+    auto EORDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal ^ Desired;
+    };
+
+    auto NEGDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return -SrcVal;
+    };
+
+    auto SWAPDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return Desired;
+    };
+
+    CASDesiredFn<AtomicType> DesiredFunction{};
+
+    switch (AtomicOp) {
+      case ExclusiveAtomicPairType::TYPE_SWAP:
+        DesiredFunction = SWAPDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_ADD:
+        DesiredFunction = ADDDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_SUB:
+        DesiredFunction = SUBDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_AND:
+        DesiredFunction = ANDDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_OR:
+        DesiredFunction = ORDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_EOR:
+        DesiredFunction = EORDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_NEG:
+        DesiredFunction = NEGDesired;
+        break;
+      default:
+        LogMan::Msg::E("Unhandled JIT SIGBUS Atomic mem op 0x%02x", AtomicOp);
+        return false;
+        break;
+    }
+
+    auto Res = DoCAS16<DoRetry>(
+      mcontext->regs[DataSourceReg],
+      0, // Unused
+      Addr,
+      NOPExpected,
+      DesiredFunction);
+
+    if (AtomicFetch && ResultReg != 31) {
+      // On atomic fetch then we store the resulting value back in to the loadacquire destination register
+      // We want the memory value BEFORE the ALU op
+      mcontext->regs[ResultReg] = Res;
+    }
+  }
+  else if (Size == 4) {
+    using AtomicType = uint32_t;
+    auto NOPExpected = [](AtomicType SrcVal, AtomicType) -> AtomicType {
+      return SrcVal;
+    };
+
+    auto ADDDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal + Desired;
+    };
+
+    auto SUBDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal - Desired;
+    };
+
+    auto ANDDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal & Desired;
+    };
+
+    auto ORDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal | Desired;
+    };
+
+    auto EORDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal ^ Desired;
+    };
+
+    auto NEGDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return -SrcVal;
+    };
+
+    auto SWAPDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return Desired;
+    };
+
+    CASDesiredFn<AtomicType> DesiredFunction{};
+
+    switch (AtomicOp) {
+      case ExclusiveAtomicPairType::TYPE_SWAP:
+        DesiredFunction = SWAPDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_ADD:
+        DesiredFunction = ADDDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_SUB:
+        DesiredFunction = SUBDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_AND:
+        DesiredFunction = ANDDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_OR:
+        DesiredFunction = ORDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_EOR:
+        DesiredFunction = EORDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_NEG:
+        DesiredFunction = NEGDesired;
+        break;
+      default:
+        LogMan::Msg::E("Unhandled JIT SIGBUS Atomic mem op 0x%02x", AtomicOp);
+        return false;
+        break;
+    }
+
+    auto Res = DoCAS32<DoRetry>(
+      mcontext->regs[DataSourceReg],
+      0, // Unused
+      Addr,
+      NOPExpected,
+      DesiredFunction);
+
+    if (AtomicFetch && ResultReg != 31) {
+      // On atomic fetch then we store the resulting value back in to the loadacquire destination register
+      // We want the memory value BEFORE the ALU op
+      mcontext->regs[ResultReg] = Res;
+    }
+  }
+  else if (Size == 8) {
+    using AtomicType = uint64_t;
+    auto NOPExpected = [](AtomicType SrcVal, AtomicType) -> AtomicType {
+      return SrcVal;
+    };
+
+    auto ADDDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal + Desired;
+    };
+
+    auto SUBDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal - Desired;
+    };
+
+    auto ANDDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal & Desired;
+    };
+
+    auto ORDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal | Desired;
+    };
+
+    auto EORDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return SrcVal ^ Desired;
+    };
+
+    auto NEGDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return -SrcVal;
+    };
+
+    auto SWAPDesired = [](AtomicType SrcVal, AtomicType Desired) -> AtomicType {
+      return Desired;
+    };
+
+    CASDesiredFn<AtomicType> DesiredFunction{};
+
+    switch (AtomicOp) {
+      case ExclusiveAtomicPairType::TYPE_SWAP:
+        DesiredFunction = SWAPDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_ADD:
+        DesiredFunction = ADDDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_SUB:
+        DesiredFunction = SUBDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_AND:
+        DesiredFunction = ANDDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_OR:
+        DesiredFunction = ORDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_EOR:
+        DesiredFunction = EORDesired;
+        break;
+      case ExclusiveAtomicPairType::TYPE_NEG:
+        DesiredFunction = NEGDesired;
+        break;
+      default:
+        LogMan::Msg::E("Unhandled JIT SIGBUS Atomic mem op 0x%02x", AtomicOp);
+        return false;
+        break;
+    }
+
+    auto Res = DoCAS64<DoRetry>(
+      mcontext->regs[DataSourceReg],
+      0, // Unused
+      Addr,
+      NOPExpected,
+      DesiredFunction);
+    if (AtomicFetch && ResultReg != 31) {
+      // On atomic fetch then we store the resulting value back in to the loadacquire destination register
+      // We want the memory value BEFORE the ALU op
+      mcontext->regs[ResultReg] = Res;
+    }
+  }
+
+  // Multiply by 4 for number of bytes to skip
+  return NumInstructionsToSkip * 4;
+}
+
 }

External/FEXCore/Source/Interface/Core/ArchHelpers/Arm64.h

@@ -18,6 +18,36 @@ namespace FEXCore::ArchHelpers::Arm64 {
   constexpr uint32_t STLXP_MASK = 0xBF'E0'80'00;
   constexpr uint32_t STLXP_INST = 0x88'20'80'00;
 
+  constexpr uint32_t LDAXR_MASK = 0x3F'FF'FC'00;
+  constexpr uint32_t LDAXR_INST = 0x08'5F'FC'00;
+
+  constexpr uint32_t STLXR_MASK = 0x3F'E0'FC'00;
+  constexpr uint32_t STLXR_INST = 0x08'00'FC'00;
+
+  constexpr uint32_t CBNZ_MASK = 0x7F'00'00'00;
+  constexpr uint32_t CBNZ_INST = 0x35'00'00'00;
+
+  constexpr uint32_t ALU_OP_MASK = 0x7F'00'00'00;
+  constexpr uint32_t ADD_INST    = 0x0B'00'00'00;
+  constexpr uint32_t SUB_INST    = 0x4B'00'00'00;
+  constexpr uint32_t CMP_INST    = 0x6B'00'00'00;
+  constexpr uint32_t AND_INST    = 0x0A'00'00'00;
+  constexpr uint32_t OR_INST     = 0x2A'00'00'00;
+  constexpr uint32_t EOR_INST    = 0x4A'00'00'00;
+  
+  constexpr uint32_t CCMP_MASK   = 0x7F'E0'0C'10;
+  constexpr uint32_t CCMP_INST   = 0x7A'40'00'00;
+  
+  enum ExclusiveAtomicPairType {
+    TYPE_SWAP,
+    TYPE_ADD,
+    TYPE_SUB,
+    TYPE_AND,
+    TYPE_OR,
+    TYPE_EOR,
+    TYPE_NEG, // This is just a sub with zero. Need to know the differences
+  };
+
   // Load ops are 4 bits
   // Acquire and release bits are independent on the instruction
   constexpr uint32_t ATOMIC_ADD_OP  = 0b0000;
@@ -30,7 +60,29 @@ namespace FEXCore::ArchHelpers::Arm64 {
   constexpr uint32_t ATOMIC_UMIN_OP = 0b0111;
   constexpr uint32_t ATOMIC_SWAP_OP = 0b1000;
 
+  constexpr uint32_t REGISTER_MASK = 0b11111;
+  constexpr uint32_t RD_OFFSET = 0;
+  constexpr uint32_t RN_OFFSET = 5;
+  constexpr uint32_t RM_OFFSET = 16;
+
+  inline uint32_t GetRdReg(uint32_t Instr) {
+    return (Instr >> RD_OFFSET) & REGISTER_MASK;
+  }
+
+  inline uint32_t GetRnReg(uint32_t Instr) {
+    return (Instr >> RN_OFFSET) & REGISTER_MASK;
+  }
+
+  inline uint32_t GetRmReg(uint32_t Instr) {
+    return (Instr >> RM_OFFSET) & REGISTER_MASK;
+  }
+
+  bool HandleAtomicLoad(void *_ucontext, void *_info, uint32_t Instr);
+  bool HandleAtomicStore(void *_ucontext, void *_info, uint32_t Instr);
+  bool HandleAtomicLoad128(void *_ucontext, void *_info, uint32_t Instr);
+  uint64_t HandleAtomicLoadstoreExclusive(void *_ucontext, void *_info);
   bool HandleCASPAL(void *_ucontext, void *_info, uint32_t Instr);
+  uint64_t HandleCASPAL_ARMv8(void *_ucontext, void *_info, uint32_t Instr);
   bool HandleCASAL(void *_ucontext, void *_info, uint32_t Instr);
   bool HandleAtomicMemOp(void *_ucontext, void *_info, uint32_t Instr);
 }

External/FEXCore/Source/Interface/Core/ArchHelpers/Arm64Emitter.cpp

@@ -4,6 +4,11 @@
 #include <FEXCore/Core/CoreState.h>
 
 #include "aarch64/cpu-aarch64.h"
+#include "cpu-features.h"
+#include "aarch64/instructions-aarch64.h"
+#include "utils-vixl.h"
+
+#include <tuple>
 
 namespace FEXCore::CPU {
 #define STATE x28
@@ -143,22 +148,26 @@ void Arm64Emitter::PopCalleeSavedRegisters() {
 
 
 void Arm64Emitter::SpillStaticRegs() {
-  for (size_t i = 0; i < SRA64.size(); i+=2) {
-      stp(SRA64[i], SRA64[i+1], MemOperand(STATE, offsetof(FEXCore::Core::CpuStateFrame, State.gregs[i])));
-  }
+  if (StaticRegisterAllocation()) {
+    for (size_t i = 0; i < SRA64.size(); i+=2) {
+        stp(SRA64[i], SRA64[i+1], MemOperand(STATE, offsetof(FEXCore::Core::CpuStateFrame, State.gregs[i])));
+    }
 
-  for (size_t i = 0; i < SRAFPR.size(); i+=2) {
-    stp(SRAFPR[i].Q(), SRAFPR[i+1].Q(), MemOperand(STATE, offsetof(FEXCore::Core::CpuStateFrame, State.xmm[i][0])));
+    for (size_t i = 0; i < SRAFPR.size(); i+=2) {
+      stp(SRAFPR[i].Q(), SRAFPR[i+1].Q(), MemOperand(STATE, offsetof(FEXCore::Core::CpuStateFrame, State.xmm[i][0])));
+    }
   }
 }
 
 void Arm64Emitter::FillStaticRegs() {
-  for (size_t i = 0; i < SRA64.size(); i+=2) {
-    ldp(SRA64[i], SRA64[i+1], MemOperand(STATE, offsetof(FEXCore::Core::CpuStateFrame, State.gregs[i])));
-  }
+  if (StaticRegisterAllocation()) {
+    for (size_t i = 0; i < SRA64.size(); i+=2) {
+      ldp(SRA64[i], SRA64[i+1], MemOperand(STATE, offsetof(FEXCore::Core::CpuStateFrame, State.gregs[i])));
+    }
 
-  for (size_t i = 0; i < SRAFPR.size(); i+=2) {
-    ldp(SRAFPR[i].Q(), SRAFPR[i+1].Q(), MemOperand(STATE, offsetof(FEXCore::Core::CpuStateFrame, State.xmm[i][0])));
+    for (size_t i = 0; i < SRAFPR.size(); i+=2) {
+      ldp(SRAFPR[i].Q(), SRAFPR[i+1].Q(), MemOperand(STATE, offsetof(FEXCore::Core::CpuStateFrame, State.xmm[i][0])));
+    }
   }
 }
 
@@ -222,7 +231,7 @@ void Arm64Emitter::ResetStack() {
 }
 
 void Arm64Emitter::Align16B() {
-  uint64_t CurrentOffset = GetBuffer()->GetOffsetAddress<uint64_t>(GetCursorOffset());
+  uint64_t CurrentOffset = GetCursorAddress<uint64_t>();
   for (uint64_t i = (16 - (CurrentOffset & 0xF)); i != 0; i -= 4) {
       nop();
   }

External/FEXCore/Source/Interface/Core/ArchHelpers/Arm64Emitter.h

@@ -1,7 +1,16 @@
 #pragma once
 
 #include "aarch64/assembler-aarch64.h"
+#include "aarch64/constants-aarch64.h"
 #include "aarch64/cpu-aarch64.h"
+#include "aarch64/operands-aarch64.h"
+#include "platform-vixl.h"
+#include "FEXCore/Config/Config.h"
+
+#include <array>
+#include <stddef.h>
+#include <stdint.h>
+#include <utility>
 
 namespace FEXCore::CPU {
 using namespace vixl;
@@ -72,6 +81,8 @@ protected:
 
   uint32_t DCacheLineSize{};
   uint32_t ICacheLineSize{};
+
+  FEX_CONFIG_OPT(StaticRegisterAllocation, SRA);
 };
 
 }

External/FEXCore/Source/Interface/Core/ArchHelpers/Arm64_stubs.cpp

@@ -1,6 +1,7 @@
 #include "Interface/Core/ArchHelpers/Arm64.h"
 
 #include <FEXCore/Utils/LogManager.h>
+#include <stdint.h>
 
 namespace FEXCore::ArchHelpers::Arm64 {
 
@@ -23,4 +24,4 @@ bool HandleAtomicMemOp(void *_ucontext, void *_info, uint32_t Instr) {
 }
 #endif
 
-}
+}

External/FEXCore/Source/Interface/Core/ArchHelpers/MContext.h

@@ -18,6 +18,7 @@ struct X86ContextBackup {
   // RIP and RSP is stored in GPRs here
   uint64_t GPRs[23];
   FEXCore::x86_64::_libc_fpstate FPRState;
+  uint64_t sa_mask;
 
   // Guest state
   int Signal;
@@ -35,6 +36,7 @@ struct ArmContextBackup {
   uint32_t FPSR;
   uint32_t FPCR;
   __uint128_t FPRs[32];
+  uint64_t sa_mask;
 
   // Guest state
   int Signal;
@@ -44,6 +46,11 @@ struct ArmContextBackup {
   static constexpr int RedZoneSize = 0;
 };
 
+static inline ucontext_t* GetUContext(void* ucontext) {
+  ucontext_t* _context = (ucontext_t*)ucontext;
+  return _context;
+}
+
 static inline mcontext_t* GetMContext(void* ucontext) {
   ucontext_t* _context = (ucontext_t*)ucontext;
   return &_context->uc_mcontext;
@@ -102,6 +109,7 @@ using ContextBackup = ArmContextBackup;
 template <typename T>
 static inline void BackupContext(void* ucontext, T *Backup) {
   if constexpr (std::is_same<T, ArmContextBackup>::value) {
+    auto _ucontext = GetUContext(ucontext);
     auto _mcontext = GetMContext(ucontext);
 
     memcpy(&Backup->GPRs[0], &_mcontext->regs[0], 31 * sizeof(uint64_t));
@@ -115,6 +123,9 @@ static inline void BackupContext(void* ucontext, T *Backup) {
     Backup->FPSR = HostState->FPSR;
     Backup->FPCR = HostState->FPCR;
     memcpy(&Backup->FPRs[0], &HostState->FPRs[0], 32 * sizeof(__uint128_t));
+
+    // Save the signal mask so we can restore it
+    memcpy(&Backup->sa_mask, &_ucontext->uc_sigmask, sizeof(uint64_t));
   } else {
       ERROR_AND_DIE("Wrong context type"); // This must be a runtime error
   }
@@ -123,6 +134,7 @@ static inline void BackupContext(void* ucontext, T *Backup) {
 template <typename T>
 static inline void RestoreContext(void* ucontext, T *Backup) {
   if constexpr (std::is_same<T, ArmContextBackup>::value) {
+    auto _ucontext = GetUContext(ucontext);
    auto _mcontext = GetMContext(ucontext);
 
     HostFPRState *HostState = reinterpret_cast<HostFPRState*>(&_mcontext->__reserved[0]);
@@ -136,6 +148,9 @@ static inline void RestoreContext(void* ucontext, T *Backup) {
     ArchHelpers::Context::SetPc(ucontext, Backup->PrevPC);
     ArchHelpers::Context::SetSp(ucontext, Backup->PrevSP);
     memcpy(&_mcontext->regs[0], &Backup->GPRs[0], 31 * sizeof(uint64_t));
+
+    // Restore the signal mask now
+    memcpy(&_ucontext->uc_sigmask, &Backup->sa_mask, sizeof(uint64_t));
   } else {
       ERROR_AND_DIE("Wrong context type"); // This must be a runtime error
   }
@@ -181,6 +196,7 @@ using ContextBackup = X86ContextBackup;
 template <typename T>
 static inline void BackupContext(void* ucontext, T *Backup) {
   if constexpr (std::is_same<T, X86ContextBackup>::value) {
+    auto _ucontext = GetUContext(ucontext);
     auto _mcontext = GetMContext(ucontext);
 
     // Copy the GPRs
@@ -188,6 +204,9 @@ static inline void BackupContext(void* ucontext, T *Backup) {
     // Copy the FPRState
     memcpy(&Backup->FPRState, _mcontext->fpregs, sizeof(X86ContextBackup::FPRState));
     // XXX: Save 256bit and 512bit AVX register state
+
+    // Save the signal mask so we can restore it
+    memcpy(&Backup->sa_mask, &_ucontext->uc_sigmask, sizeof(uint64_t));
   } else {
     ERROR_AND_DIE("Wrong context type"); // This must be a runtime error
   }
@@ -196,12 +215,16 @@ static inline void BackupContext(void* ucontext, T *Backup) {
 template <typename T>
 static inline void RestoreContext(void* ucontext, T *Backup) {
   if constexpr (std::is_same<T, X86ContextBackup>::value) {
+    auto _ucontext = GetUContext(ucontext);
     auto _mcontext = GetMContext(ucontext);
 
     // Copy the GPRs
     memcpy(&_mcontext->gregs[0], &Backup->GPRs[0], sizeof(X86ContextBackup::GPRs));
     // Copy the FPRState
     memcpy(_mcontext->fpregs, &Backup->FPRState, sizeof(X86ContextBackup::FPRState));
+
+    // Restore the signal mask now
+    memcpy(&_ucontext->uc_sigmask, &Backup->sa_mask, sizeof(uint64_t));
   } else {
     ERROR_AND_DIE("Wrong context type"); // This must be a runtime error
   }
@@ -209,4 +232,4 @@ static inline void RestoreContext(void* ucontext, T *Backup) {
 
 #endif
 
-} // namespace FEXCore::ArchHelpers::Context
+} // namespace FEXCore::ArchHelpers::Context

External/FEXCore/Source/Interface/Core/BlockSamplingData.cpp

@@ -2,6 +2,7 @@
 #include <FEXCore/Utils/LogManager.h>
 #include <cstring>
 #include <fstream>
+#include <utility>
 
 namespace FEXCore {
   void BlockSamplingData::DumpBlockData() {

External/FEXCore/Source/Interface/Core/BlockSamplingData.h

@@ -1,6 +1,7 @@
 #pragma once
+
+#include <cstdint>
 #include <unordered_map>
-#include <stdint.h>
 
 namespace FEXCore {
 class BlockSamplingData {

External/FEXCore/Source/Interface/Core/CPUID.cpp

@@ -5,8 +5,11 @@ desc: Handles presented capability bits for guest cpu
 $end_info$
 */
 
+#include <FEXCore/Config/Config.h>
+#include <FEXCore/Core/CPUID.h>
 #include "Interface/Context/Context.h"
 #include "Interface/Core/CPUID.h"
+#include "Interface/Core/HostFeatures.h"
 #include "git_version.h"
 
 #include <cstring>
@@ -96,7 +99,7 @@ FEXCore::CPUID::FunctionResults CPUIDEmu::Function_01h(uint32_t Leaf) {
     (0 <<  5) | // VMX
     (0 <<  6) | // SMX
     (0 <<  7) | // Intel SpeedStep
-    (0 <<  8) | // Thermal Monitor 2
+    (1 <<  8) | // Thermal Monitor 2
     (1 <<  9) | // SSSE3
     (0 << 10) | // L1 context ID
     (0 << 11) | // Silicon debug
@@ -151,7 +154,7 @@ FEXCore::CPUID::FunctionResults CPUIDEmu::Function_01h(uint32_t Leaf) {
     (1 << 26) | // SSE2
     (0 << 27) | // Self Snoop
     (1 << 28) | // Max APIC IDs reserved field is valid
-    (0 << 29) | // Thermal monitor
+    (1 << 29) | // Thermal monitor
     (0 << 30) | // Reserved
     (0 << 31);  // Pending break enable
   return Res;

External/FEXCore/Source/Interface/Core/CPUID.h

@@ -4,7 +4,9 @@
 
 #include <FEXCore/Core/CPUID.h>
 #include <FEXCore/Config/Config.h>
-#include <FEXCore/Utils/LogManager.h>
+
+#include <cstdint>
+#include <utility>
 
 namespace FEXCore {
 namespace Context {
@@ -28,9 +30,6 @@ public:
     auto Handler = FunctionHandlers.find(Function);
 
     if (Handler == FunctionHandlers.end()) {
-      #ifndef NDEBUG
-        LogMan::Msg::E("Unhandled CPU ID function, 0x%x-0x%x", Function, Leaf);
-      #endif
       return Function_Reserved(Leaf);
     }
 

External/FEXCore/Source/Interface/Core/CompileService.cpp

@@ -1,8 +1,21 @@
 #include "Interface/Context/Context.h"
 #include "Interface/Core/LookupCache.h"
 #include "Interface/Core/CompileService.h"
-#include "Interface/Core/InternalThreadState.h"
 #include "Interface/Core/OpcodeDispatcher.h"
+#include "FEXCore/Debug/InternalThreadState.h"
+#include "FEXCore/HLE/Linux/ThreadManagement.h"
+#include "Interface/IR/PassManager.h"
+
+#include <FEXCore/Core/CPUBackend.h>
+#include <FEXCore/Core/CoreState.h>
+#include <FEXCore/Core/SignalDelegator.h>
+#include <FEXCore/Utils/Event.h>
+#include <FEXCore/Utils/LogManager.h>
+#include <FEXCore/Utils/Threads.h>
+
+#include <memory>
+#include <pthread.h>
+#include <stdio.h>
 
 namespace FEXCore {
   static void* ThreadHandler(void *Arg) {

External/FEXCore/Source/Interface/Core/CompileService.h

@@ -1,23 +1,22 @@
 #pragma once
 
-#include <FEXCore/Core/CPUBackend.h>
+#include <FEXCore/Debug/InternalThreadState.h>
 #include <FEXCore/Utils/Event.h>
 #include <FEXCore/Utils/Threads.h>
 
+#include <atomic>
 #include <memory>
-#include <thread>
-#include <unordered_map>
+#include <mutex>
 #include <queue>
+#include <stdint.h>
+#include <vector>
 
 namespace FEXCore {
 namespace Context {
   struct Context;
 }
-namespace Core {
-  struct InternalThreadState;
-}
-
 namespace IR {
+  class IRListView;
   class RegisterAllocationData;
 };
 class CompileService final {

External/FEXCore/Source/Interface/Core/Core.cpp

@@ -7,42 +7,70 @@ desc: Glues Frontend, OpDispatcher and IR Opts & Compilation, LookupCache, Dispa
 $end_info$
 */
 
-#include "Common/MathUtils.h"
-#include "Common/Paths.h"
-
 #include "Interface/Context/Context.h"
 #include "Interface/Core/LookupCache.h"
-#include "Interface/Core/BlockSamplingData.h"
 #include "Interface/Core/CompileService.h"
 #include "Interface/Core/Core.h"
-#include "Interface/Core/DebugData.h"
+#include "Interface/Core/CPUID.h"
+#include "Interface/Core/Frontend.h"
+#include "Interface/Core/GdbServer.h"
 #include "Interface/Core/OpcodeDispatcher.h"
 #include "Interface/Core/Interpreter/InterpreterCore.h"
 #include "Interface/Core/JIT/JITCore.h"
+#include "Interface/HLE/Thunks/Thunks.h"
 #include "Interface/IR/Passes/RegisterAllocationPass.h"
 #include "Interface/IR/Passes.h"
+#include "Interface/IR/PassManager.h"
 
 #include <FEXCore/Config/Config.h>
 #include <FEXCore/Core/CodeLoader.h>
+#include <FEXCore/Core/Context.h>
 #include <FEXCore/Core/CoreState.h>
 #include <FEXCore/Core/CPUBackend.h>
+#include <FEXCore/Core/SignalDelegator.h>
 #include <FEXCore/Core/X86Enums.h>
+#include <FEXCore/Debug/InternalThreadState.h>
+#include <FEXCore/Debug/X86Tables.h>
 #include <FEXCore/HLE/SyscallHandler.h>
+#include <FEXCore/HLE/Linux/ThreadManagement.h>
+#include <FEXCore/IR/IR.h>
+#include <FEXCore/IR/IREmitter.h>
+#include <FEXCore/IR/IntrusiveIRList.h>
+#include <FEXCore/IR/RegisterAllocationData.h>
 #include <FEXCore/Utils/Allocator.h>
+#include <FEXCore/Utils/Event.h>
+#include <FEXCore/Utils/LogManager.h>
+#include <FEXCore/Utils/Threads.h>
 
-#include "Interface/HLE/Thunks/Thunks.h"
-#include "FEXCore/Utils/Allocator.h"
-
-#include <xxh3.h>
-#include <fstream>
-#include <unistd.h>
-#include <filesystem>
 #include <algorithm>
+#include <array>
+#include <atomic>
+#include <chrono>
+#include <condition_variable>
+#include <cstdint>
+#include <filesystem>
+#include <functional>
+#include <map>
+#include <memory>
+#include <mutex>
+#include <queue>
+#include <set>
+#include <shared_mutex>
+#include <signal.h>
+#include <stdio.h>
+#include <string.h>
+#include <string>
+#include <string_view>
+#include <sstream>
 #include <sys/mman.h>
-#include <unistd.h>
 #include <sys/stat.h>
-
-#include "Interface/Core/GdbServer.h"
+#include <sys/syscall.h>
+#include <type_traits>
+#include <unistd.h>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+#include <xxhash.h>
 
 namespace FEXCore::CPU {
   bool CreateCPUCore(FEXCore::Context::Context *CTX) {
@@ -195,11 +223,7 @@ namespace FEXCore::Context {
     }
   }
 
-  bool Context::InitCore(FEXCore::CodeLoader *Loader) {
-    ThunkHandler.reset(FEXCore::ThunkHandler::Create());
-
-    LocalLoader = Loader;
-    using namespace FEXCore::Core;
+  static FEXCore::Core::CPUState CreateDefaultCPUState() {
     FEXCore::Core::CPUState NewThreadState{};
 
     // Initialize default CPU state
@@ -216,7 +240,16 @@ namespace FEXCore::Context {
     NewThreadState.flags[1] = 1;
     NewThreadState.flags[9] = 1;
     NewThreadState.FCW = 0x37F;
+    NewThreadState.FTW = 0xFFFF;
+    return NewThreadState;
+  }
 
+  FEXCore::Core::InternalThreadState* Context::InitCore(FEXCore::CodeLoader *Loader) {
+    ThunkHandler.reset(FEXCore::ThunkHandler::Create());
+
+    LocalLoader = Loader;
+    using namespace FEXCore::Core;
+    FEXCore::Core::CPUState NewThreadState = CreateDefaultCPUState();
     FEXCore::Core::InternalThreadState *Thread = CreateThread(&NewThreadState, 0);
 
     // We are the parent thread
@@ -227,8 +260,7 @@ namespace FEXCore::Context {
     Thread->CurrentFrame->State.rip = StartingRIP = Loader->DefaultRIP();
 
     InitializeThreadData(Thread);
-
-    return true;
+    return Thread;
   }
 
   void Context::StartGdbServer() {
@@ -242,8 +274,7 @@ namespace FEXCore::Context {
     DebugServer.reset();
   }
 
-  void Context::HandleCallback(uint64_t RIP) {
-    auto Thread = Core::ThreadData.Thread;
+  void Context::HandleCallback(FEXCore::Core::InternalThreadState *Thread, uint64_t RIP) {
     Thread->CPUBackend->CallbackPtr(Thread->CurrentFrame, RIP);
   }
 
@@ -424,7 +455,13 @@ namespace FEXCore::Context {
     auto IRHandler = [Thread](uint64_t Addr, IR::IREmitter *IR) -> void {
       // Run the passmanager over the IR from the dispatcher
       Thread->PassManager->Run(IR);
-      Core::LocalIREntry Entry = {Addr, 0ULL, decltype(Entry.IR)(IR->CreateIRCopy()), decltype(Entry.RAData)(Thread->PassManager->GetRAPass() ? Thread->PassManager->GetRAPass()->PullAllocationData() : nullptr), decltype(Entry.DebugData)(new Core::DebugData())};
+      Core::LocalIREntry Entry = {Addr, 0ULL,
+        decltype(Entry.IR)(IR->CreateIRCopy()),
+        decltype(Entry.RAData)(Thread->PassManager->HasPass("RA")
+          ? Thread->PassManager->GetPass<IR::RegisterAllocationPass>("RA")->PullAllocationData()
+          : nullptr),
+        decltype(Entry.DebugData)(new Core::DebugData())
+      };
       Thread->LocalIRCache.insert({Addr, std::move(Entry)});
     };
 
@@ -455,6 +492,14 @@ namespace FEXCore::Context {
     Thread->ThreadWaiting.Wait();
   }
 
+  void Context::InitializeThreadTLSData(FEXCore::Core::InternalThreadState *Thread) {
+    // Let's do some initial bookkeeping here
+    Thread->ThreadManager.TID = ::gettid();
+    Thread->ThreadManager.PID = ::getpid();
+    SignalDelegation->RegisterTLSState(Thread);
+    ThunkHandler->RegisterTLSState(Thread);
+  }
+
   void Context::RunThread(FEXCore::Core::InternalThreadState *Thread) {
     // Tell the thread to start executing
     Thread->StartRunning.NotifyAll();
@@ -470,8 +515,10 @@ namespace FEXCore::Context {
         Stop(false /* Ignore current thread */);
     });
 
+    State->CTX = this;
+
     #if _M_ARM_64
-    bool DoSRA = true;
+    bool DoSRA = State->CTX->Config.StaticRegisterAllocation;
     #else
     bool DoSRA = false;
     #endif
@@ -481,8 +528,6 @@ namespace FEXCore::Context {
 
     State->PassManager->RegisterSyscallHandler(SyscallHandler);
 
-    State->CTX = this;
-
     // Create CPU backend
     switch (Config.Core) {
     case FEXCore::Config::CONFIG_INTERPRETER:
@@ -586,6 +631,12 @@ namespace FEXCore::Context {
 
     // Clean up dead stacks
     FEXCore::Threads::Thread::CleanupAfterFork();
+
+    if (LiveThread->CompileService) {
+      // If this live thread had a compile service then it no longer exists
+      // Erase the shared_ptr
+      LiveThread->CompileService.reset();
+    }
   }
 
   void Context::AddBlockMapping(FEXCore::Core::InternalThreadState *Thread, uint64_t Address, void *Ptr, uint64_t Start, uint64_t Length) {
@@ -777,17 +828,17 @@ namespace FEXCore::Context {
     Thread->PassManager->Run(Thread->OpDispatcher.get());
 
     if (Thread->CTX->Config.DumpIR() != "no") {
-      IRDumper(Thread->PassManager->GetRAPass() ? Thread->PassManager->GetRAPass()->GetAllocationData() : nullptr);
+      IRDumper(Thread->PassManager->HasPass("RA") ? Thread->PassManager->GetPass<IR::RegisterAllocationPass>("RA")->GetAllocationData() : nullptr);
     }
 
     if (Thread->OpDispatcher->ShouldDump) {
       std::stringstream out;
       auto NewIR = Thread->OpDispatcher->ViewIR();
-      FEXCore::IR::Dump(&out, &NewIR, Thread->PassManager->GetRAPass() ? Thread->PassManager->GetRAPass()->GetAllocationData() : nullptr);
+      FEXCore::IR::Dump(&out, &NewIR, Thread->PassManager->HasPass("RA") ? Thread->PassManager->GetPass<IR::RegisterAllocationPass>("RA")->GetAllocationData() : nullptr);
       LogMan::Msg::I("IR 0x%lx:\n%s\n@@@@@\n", GuestRIP, out.str().c_str());
     }
 
-    auto RAData = Thread->PassManager->GetRAPass() ? Thread->PassManager->GetRAPass()->PullAllocationData() : nullptr;
+    auto RAData = Thread->PassManager->HasPass("RA") ? Thread->PassManager->GetPass<IR::RegisterAllocationPass>("RA")->PullAllocationData() : nullptr;
     auto IRList = Thread->OpDispatcher->CreateIRCopy();
 
     Thread->OpDispatcher->ResetWorkingList();
@@ -1084,7 +1135,9 @@ namespace FEXCore::Context {
 
     if (NewBlock == 0) {
       LogMan::Msg::E("CompileBlockJit: Failed to compile code %lX - aborting process", GuestRIP);
-      abort();
+      // Return similar behaviour of SIGILL abort
+      Frame->Thread->StatusCode = 128 + SIGILL;
+      Stop(false /* Ignore current thread */);
     }
   }
 
@@ -1183,8 +1236,6 @@ namespace FEXCore::Context {
                 AotFile->Stream->write((char*)&tag, sizeof(tag));
               }
               AotFile->AppendAOTIRCaptureCache(LocalRIP, LocalStartAddr, Length, hash, IRList, RAData);
-              delete IRList;
-              FEXCore::Allocator::free(RAData);
             });
           }
         }
@@ -1219,11 +1270,7 @@ namespace FEXCore::Context {
     Core::ThreadData.Thread = Thread;
     Thread->ExitReason = FEXCore::Context::ExitReason::EXIT_WAITING;
 
-    // Let's do some initial bookkeeping here
-    Thread->ThreadManager.TID = ::gettid();
-    Thread->ThreadManager.PID = ::getpid();
-    SignalDelegation->RegisterTLSState(Thread);
-    ThunkHandler->RegisterTLSState(Thread);
+    InitializeThreadTLSData(Thread);
 
     ++IdleWaitRefCount;
 

External/FEXCore/Source/Interface/Core/Dispatcher/Arm64Dispatcher.cpp

@@ -1,16 +1,33 @@
+#include "Interface/Core/LookupCache.h"
+
 #include "Interface/Core/ArchHelpers/MContext.h"
 #include "Interface/Core/Dispatcher/Arm64Dispatcher.h"
-
 #include "Interface/Core/Interpreter/InterpreterClass.h"
-#include "Interface/Context/Context.h"
 
+#include "Interface/Context/Context.h"
+#include "Interface/Core/X86HelperGen.h"
+
+#include <FEXCore/Core/CPUBackend.h>
+#include <FEXCore/Core/CoreState.h>
 #include <FEXCore/Core/X86Enums.h>
+#include <FEXCore/Debug/InternalThreadState.h>
+#include <array>
 #include <bit>
 #include <cmath>
+#include <cstdint>
+#include <memory>
+#include <stddef.h>
 
 #include "aarch64/assembler-aarch64.h"
+#include "aarch64/constants-aarch64.h"
+#include "aarch64/operands-aarch64.h"
 #include "aarch64/cpu-aarch64.h"
-#include "aarch64/disasm-aarch64.h"
+#include "code-buffer-vixl.h"
+#include "platform-vixl.h"
+
+#ifdef ENABLE_JITSYMBOLS
+#include <unistd.h>
+#endif
 
 namespace FEXCore::CPU {
 
@@ -25,8 +42,7 @@ Arm64Dispatcher::Arm64Dispatcher(FEXCore::Context::Context *ctx, FEXCore::Core::
   SRAEnabled = config.StaticRegisterAssignment;
   SetAllowAssembler(true);
 
-  auto Buffer = GetBuffer();
-  DispatchPtr = Buffer->GetOffsetAddress<CPUBackend::AsmDispatch>(GetCursorOffset());
+  DispatchPtr = GetCursorAddress<CPUBackend::AsmDispatch>();
 
   // while (true) {
   //    Ptr = FindBlock(RIP)
@@ -59,7 +75,7 @@ Arm64Dispatcher::Arm64Dispatcher(FEXCore::Context::Context *ctx, FEXCore::Core::
   add(x0, sp, 0);
   str(x0, MemOperand(STATE, offsetof(FEXCore::Core::CpuStateFrame, ReturningStackLocation)));
 
-  AbsoluteLoopTopAddressFillSRA = Buffer->GetOffsetAddress<uint64_t>(GetCursorOffset());
+  AbsoluteLoopTopAddressFillSRA = GetCursorAddress<uint64_t>();
 
   if (SRAEnabled) {
     FillStaticRegs();
@@ -178,11 +194,11 @@ Arm64Dispatcher::Arm64Dispatcher(FEXCore::Context::Context *ctx, FEXCore::Core::
 
   {
     bind(&ExitSpillSRA);
-    ThreadStopHandlerAddressSpillSRA = Buffer->GetOffsetAddress<uint64_t>(GetCursorOffset());
+    ThreadStopHandlerAddressSpillSRA = GetCursorAddress<uint64_t>();
     if (SRAEnabled)
       SpillStaticRegs();
 
-    ThreadStopHandlerAddress = Buffer->GetOffsetAddress<uint64_t>(GetCursorOffset());
+    ThreadStopHandlerAddress = GetCursorAddress<uint64_t>();
 
     PopCalleeSavedRegisters();
 
@@ -192,7 +208,7 @@ Arm64Dispatcher::Arm64Dispatcher(FEXCore::Context::Context *ctx, FEXCore::Core::
   }
 
   {
-    ExitFunctionLinkerAddress = Buffer->GetOffsetAddress<uint64_t>(GetCursorOffset());
+    ExitFunctionLinkerAddress = GetCursorAddress<uint64_t>();
     if (SRAEnabled)
       SpillStaticRegs();
 
@@ -229,7 +245,7 @@ Arm64Dispatcher::Arm64Dispatcher(FEXCore::Context::Context *ctx, FEXCore::Core::
   }
 
   {
-    SignalHandlerReturnAddress = Buffer->GetOffsetAddress<uint64_t>(GetCursorOffset());
+    SignalHandlerReturnAddress = GetCursorAddress<uint64_t>();
 
     // Now to get back to our old location we need to do a fault dance
     // We can't use SIGTRAP here since gdb catches it and never gives it to the application!
@@ -237,12 +253,12 @@ Arm64Dispatcher::Arm64Dispatcher(FEXCore::Context::Context *ctx, FEXCore::Core::
   }
 
   {
-    ThreadPauseHandlerAddressSpillSRA = Buffer->GetOffsetAddress<uint64_t>(GetCursorOffset());
+    ThreadPauseHandlerAddressSpillSRA = GetCursorAddress<uint64_t>();
     if (SRAEnabled)
       SpillStaticRegs();
 
     bind(&ThreadPauseHandler);
-    ThreadPauseHandlerAddress = Buffer->GetOffsetAddress<uint64_t>(GetCursorOffset());
+    ThreadPauseHandlerAddress = GetCursorAddress<uint64_t>();
     // We are pausing, this means the frontend should be waiting for this thread to idle
     // We will have faulted and jumped to this location at this point
 
@@ -252,7 +268,7 @@ Arm64Dispatcher::Arm64Dispatcher(FEXCore::Context::Context *ctx, FEXCore::Core::
     ldr(x2, &l_Sleep);
     blr(x2);
 
-    PauseReturnInstruction = Buffer->GetOffsetAddress<uint64_t>(GetCursorOffset());
+    PauseReturnInstruction = GetCursorAddress<uint64_t>();
     // Fault to start running again
     hlt(0);
   }
@@ -273,7 +289,7 @@ Arm64Dispatcher::Arm64Dispatcher(FEXCore::Context::Context *ctx, FEXCore::Core::
     // On return to the thunk, the thunk can get whatever its return value is from the thread context depending on ABI handling on its end
     // When the thunk itself returns, it'll do its regular return logic there
     // void ReentrantCallback(FEXCore::Core::InternalThreadState *Thread, uint64_t RIP);
-    CallbackPtr = Buffer->GetOffsetAddress<CPUBackend::JITCallback>(GetCursorOffset());
+    CallbackPtr = GetCursorAddress<CPUBackend::JITCallback>();
 
     // We expect the thunk to have previously pushed the registers it was using
     PushCalleeSavedRegisters();
@@ -322,7 +338,7 @@ Arm64Dispatcher::Arm64Dispatcher(FEXCore::Context::Context *ctx, FEXCore::Core::
 
   FinalizeCode();
   Start = reinterpret_cast<uint64_t>(DispatchPtr);
-  End = Buffer->GetOffsetAddress<uint64_t>(GetCursorOffset());
+  End = GetCursorAddress<uint64_t>();
   vixl::aarch64::CPU::EnsureIAndDCacheCoherency(reinterpret_cast<void*>(DispatchPtr), End - reinterpret_cast<uint64_t>(DispatchPtr));
   GetBuffer()->SetExecutable();
 

External/FEXCore/Source/Interface/Core/Dispatcher/Arm64Dispatcher.h

@@ -3,7 +3,13 @@
 #include "Interface/Core/ArchHelpers/Arm64Emitter.h"
 #include "Interface/Core/Dispatcher/Dispatcher.h"
 
-#include "aarch64/assembler-aarch64.h"
+namespace FEXCore::Context {
+struct Context;
+}
+
+namespace FEXCore::Core {
+struct InternalThreadState;
+}
 
 namespace FEXCore::CPU {
 
@@ -15,4 +21,4 @@ class Arm64Dispatcher final : public Dispatcher, public Arm64Emitter {
     void SpillSRA(void *ucontext) override;
 };
 
-}
+}

External/FEXCore/Source/Interface/Core/Dispatcher/Dispatcher.cpp

@@ -1,8 +1,23 @@
-#include "Interface/Core/Dispatcher/Dispatcher.h"
 
 #include "Common/MathUtils.h"
 #include "Interface/Core/ArchHelpers/MContext.h"
+#include "Interface/Core/Dispatcher/Dispatcher.h"
+#include "Interface/Core/X86HelperGen.h"
+
+#include <FEXCore/Config/Config.h>
+#include <FEXCore/Core/CoreState.h>
+#include <FEXCore/Core/SignalDelegator.h>
+#include <FEXCore/Core/UContext.h>
 #include <FEXCore/Core/X86Enums.h>
+#include <FEXCore/Debug/InternalThreadState.h>
+#include <FEXCore/Utils/Event.h>
+#include <FEXCore/Utils/LogManager.h>
+
+#include <atomic>
+#include <condition_variable>
+#include <bits/types/siginfo_t.h>
+#include <signal.h>
+#include <string.h>
 
 namespace FEXCore::CPU {
 
@@ -54,6 +69,7 @@ void Dispatcher::StoreThreadState(int Signal, void *ucontext) {
 }
 
 void Dispatcher::RestoreThreadState(void *ucontext) {
+  LOGMAN_THROW_A(!SignalFrames.empty(), "Trying to restore a signal frame when we don't have any");
   uint64_t OldSP = SignalFrames.top();
   SignalFrames.pop();
   uintptr_t NewSP = OldSP;
@@ -64,10 +80,38 @@ void Dispatcher::RestoreThreadState(void *ucontext) {
 
   // Now restore host state
   ArchHelpers::Context::RestoreContext(ucontext, Context);
+}
 
-  // Restore the previous signal state
-  // This allows recursive signals to properly handle signal masking as we are walking back up the list of signals
-  CTX->SignalDelegation->SetCurrentSignal(Context->Signal);
+static uint32_t ConvertSignalToTrapNo(int Signal, siginfo_t *HostSigInfo) {
+  switch (Signal) {
+    case SIGSEGV:
+      if (HostSigInfo->si_code == SEGV_MAPERR ||
+          HostSigInfo->si_code == SEGV_ACCERR) {
+        // Protection fault
+        return X86State::X86_TRAPNO_PF;
+      }
+      break;
+  }
+
+  // Unknown mapping, fall back to old behaviour and just pass signal
+  return Signal;
+}
+
+static uint32_t ConvertSignalToError(int Signal, siginfo_t *HostSigInfo) {
+  switch (Signal) {
+    case SIGSEGV:
+      if (HostSigInfo->si_code == SEGV_MAPERR ||
+          HostSigInfo->si_code == SEGV_ACCERR) {
+        // Protection fault
+        // Always a user fault for us
+        // XXX: PF_PROT and PF_WRITE
+        return X86State::X86_PF_USER;
+      }
+      break;
+  }
+
+  // Not a page fault issue
+  return 0;
 }
 
 bool Dispatcher::HandleGuestSignal(int Signal, void *info, void *ucontext, GuestSigAction *GuestAction, stack_t *GuestStack) {
@@ -78,60 +122,97 @@ bool Dispatcher::HandleGuestSignal(int Signal, void *info, void *ucontext, Guest
   // We use this to track if it is safe to clear cache
   ++SignalHandlerRefCounter;
 
+  uint64_t OldPC = ArchHelpers::Context::GetPc(ucontext);
   // Set the new PC
   ArchHelpers::Context::SetPc(ucontext, AbsoluteLoopTopAddressFillSRA);
   // Set our state register to point to our guest thread data
   ArchHelpers::Context::SetState(ucontext, reinterpret_cast<uint64_t>(Frame));
 
-
   uint64_t OldGuestSP = Frame->State.gregs[X86State::REG_RSP];
   uint64_t NewGuestSP = OldGuestSP;
 
-  if (!(GuestStack->ss_flags & SS_DISABLE)) {
-    // If our guest is already inside of the alternative stack
-    // Then that means we are hitting recursive signals and we need to walk back the stack correctly
-    uint64_t AltStackBase = reinterpret_cast<uint64_t>(GuestStack->ss_sp);
-    uint64_t AltStackEnd = AltStackBase + GuestStack->ss_size;
-    if (OldGuestSP >= AltStackBase &&
-        OldGuestSP <= AltStackEnd) {
-      // We are already in the alt stack, the rest of the code will handle adjusting this
-    }
-    else {
-      NewGuestSP = AltStackEnd;
+  // Pulling from context here
+  bool Is64BitMode = CTX->Config.Is64BitMode;
+  uint64_t SignalReturn = CTX->X86CodeGen.SignalReturn;
+
+  // Spill the SRA regardless of signal handler type
+  // We are going to be returning to the top of the dispatcher which will fill again
+  // Otherwise we might load garbage
+  if (SRAEnabled) {
+    if (IsAddressInJITCode(OldPC, false)) {
+      // We are in jit, SRA must be spilled
+      SpillSRA(ucontext);
+    } else {
+      if (!IsAddressInJITCode(OldPC, true)) {
+        // This is likely to cause issues but in some cases it isn't fatal
+        // This can also happen if we have put a signal on hold, then we just reenabled the signal
+        // So we are in the syscall handler
+        // Only throw a log message in this case
+        LogMan::Msg::E("Signals in dispatcher have unsynchronized context");
+      }
     }
   }
 
-  // Back up past the redzone, which is 128bytes
-  // Don't need this offset if we aren't going to be putting siginfo in to it
-  NewGuestSP -= 128;
-
-  if (GuestAction->sa_flags & SA_SIGINFO) {
-    if (SRAEnabled) {
-      if (!IsAddressInJITCode(ArchHelpers::Context::GetPc(ucontext), false)) {
-        LOGMAN_THROW_A(!IsAddressInJITCode(ArchHelpers::Context::GetPc(ucontext), true), "Signals in dispatcher have unsynchronized context");
-      } else {
-        // We are in jit, SRA must be spilled
-        SpillSRA(ucontext);
+  // altstack is only used if the signal handler was setup with SA_ONSTACK
+  if (GuestAction->sa_flags & SA_ONSTACK) {
+    // Additionally the altstack is only used if the enabled (SS_DISABLE flag is not set)
+    if (!(GuestStack->ss_flags & SS_DISABLE)) {
+      // If our guest is already inside of the alternative stack
+      // Then that means we are hitting recursive signals and we need to walk back the stack correctly
+      uint64_t AltStackBase = reinterpret_cast<uint64_t>(GuestStack->ss_sp);
+      uint64_t AltStackEnd = AltStackBase + GuestStack->ss_size;
+      if (OldGuestSP >= AltStackBase &&
+          OldGuestSP <= AltStackEnd) {
+        // We are already in the alt stack, the rest of the code will handle adjusting this
+      }
+      else {
+        NewGuestSP = AltStackEnd;
       }
     }
+  }
+
+  if (Is64BitMode) {
+    // Back up past the redzone, which is 128bytes
+    // 32-bit doesn't have a redzone
+    NewGuestSP -= 128;
+  }
+
+  // siginfo_t
+  siginfo_t *HostSigInfo = reinterpret_cast<siginfo_t*>(info);
+
+  if (GuestAction->sa_flags & SA_SIGINFO) {
 
     // Setup ucontext a bit
-    if (CTX->Config.Is64BitMode) {
+    if (Is64BitMode) {
+      NewGuestSP -= sizeof(FEXCore::x86_64::_libc_fpstate);
+      NewGuestSP = AlignDown(NewGuestSP, alignof(FEXCore::x86_64::_libc_fpstate));
+      uint64_t FPStateLocation = NewGuestSP;
 
       NewGuestSP -= sizeof(FEXCore::x86_64::ucontext_t);
+      NewGuestSP = AlignDown(NewGuestSP, alignof(FEXCore::x86_64::ucontext_t));
       uint64_t UContextLocation = NewGuestSP;
 
       NewGuestSP -= sizeof(siginfo_t);
+      NewGuestSP = AlignDown(NewGuestSP, alignof(siginfo_t));
       uint64_t SigInfoLocation = NewGuestSP;
 
       FEXCore::x86_64::ucontext_t *guest_uctx = reinterpret_cast<FEXCore::x86_64::ucontext_t*>(UContextLocation);
       siginfo_t *guest_siginfo = reinterpret_cast<siginfo_t*>(SigInfoLocation);
 
       // We have extended float information
-      guest_uctx->uc_flags |= FEXCore::x86_64::UC_FP_XSTATE;
+      guest_uctx->uc_flags = FEXCore::x86_64::UC_FP_XSTATE;
 
       // Pointer to where the fpreg memory is
-      guest_uctx->uc_mcontext.fpregs = &guest_uctx->__fpregs_mem;
+      guest_uctx->uc_mcontext.fpregs = reinterpret_cast<FEXCore::x86_64::_libc_fpstate*>(FPStateLocation);
+      FEXCore::x86_64::_libc_fpstate *fpstate = reinterpret_cast<FEXCore::x86_64::_libc_fpstate*>(FPStateLocation);
+
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86_64::FEX_REG_RIP] = Frame->State.rip;
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86_64::FEX_REG_EFL] = 0;
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86_64::FEX_REG_CSGSFS] = 0;
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86_64::FEX_REG_ERR] = ConvertSignalToError(Signal, HostSigInfo);
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86_64::FEX_REG_TRAPNO] = ConvertSignalToTrapNo(Signal, HostSigInfo);
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86_64::FEX_REG_OLDMASK] = 0;
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86_64::FEX_REG_CR2] = 0;
 
 #define COPY_REG(x) \
       guest_uctx->uc_mcontext.gregs[FEXCore::x86_64::FEX_REG_##x] = Frame->State.gregs[X86State::REG_##x];
@@ -154,14 +235,15 @@ bool Dispatcher::HandleGuestSignal(int Signal, void *info, void *ucontext, Guest
 #undef COPY_REG
 
       // Copy float registers
-      memcpy(guest_uctx->__fpregs_mem._st, Frame->State.mm, sizeof(Frame->State.mm));
-      memcpy(guest_uctx->__fpregs_mem._xmm, Frame->State.xmm, sizeof(Frame->State.xmm));
+      memcpy(fpstate->_st, Frame->State.mm, sizeof(Frame->State.mm));
+      memcpy(fpstate->_xmm, Frame->State.xmm, sizeof(Frame->State.xmm));
 
       // FCW store default
-      guest_uctx->__fpregs_mem.fcw = Frame->State.FCW;
+      fpstate->fcw = Frame->State.FCW;
+      fpstate->ftw = Frame->State.FTW;
 
       // Reconstruct FSW
-      guest_uctx->__fpregs_mem.fsw =
+      fpstate->fsw =
         (Frame->State.flags[FEXCore::X86State::X87FLAG_TOP_LOC] << 11) |
         (Frame->State.flags[FEXCore::X86State::X87FLAG_C0_LOC] << 8) |
         (Frame->State.flags[FEXCore::X86State::X87FLAG_C1_LOC] << 9) |
@@ -173,8 +255,6 @@ bool Dispatcher::HandleGuestSignal(int Signal, void *info, void *ucontext, Guest
       guest_uctx->uc_stack.ss_sp = GuestStack->ss_sp;
       guest_uctx->uc_stack.ss_size = GuestStack->ss_size;
 
-      // siginfo_t
-      siginfo_t *HostSigInfo = reinterpret_cast<siginfo_t*>(info);
       // aarch64 and x86_64 siginfo_t matches. We can just copy this over
       // SI_USER could also potentially have random data in it, needs to be bit perfect
       // For guest faults we don't have a real way to reconstruct state to a real guest RIP
@@ -184,14 +264,77 @@ bool Dispatcher::HandleGuestSignal(int Signal, void *info, void *ucontext, Guest
       Frame->State.gregs[X86State::REG_RDX] = UContextLocation;
     }
     else {
-      // XXX: 32bit Support
+      NewGuestSP -= sizeof(FEXCore::x86::_libc_fpstate);
+      NewGuestSP = AlignDown(NewGuestSP, alignof(FEXCore::x86::_libc_fpstate));
+      uint64_t FPStateLocation = NewGuestSP;
+
       NewGuestSP -= sizeof(FEXCore::x86::ucontext_t);
-      uint64_t UContextLocation = 0; // NewGuestSP;
+      NewGuestSP = AlignDown(NewGuestSP, alignof(FEXCore::x86::ucontext_t));
+      uint64_t UContextLocation = NewGuestSP;
+
       NewGuestSP -= sizeof(FEXCore::x86::siginfo_t);
+      NewGuestSP = AlignDown(NewGuestSP, alignof(FEXCore::x86::siginfo_t));
       uint64_t SigInfoLocation = NewGuestSP;
 
+      FEXCore::x86::ucontext_t *guest_uctx = reinterpret_cast<FEXCore::x86::ucontext_t*>(UContextLocation);
       FEXCore::x86::siginfo_t *guest_siginfo = reinterpret_cast<FEXCore::x86::siginfo_t*>(SigInfoLocation);
-      siginfo_t *HostSigInfo = reinterpret_cast<siginfo_t*>(info);
+
+      // We have extended float information
+      guest_uctx->uc_flags = FEXCore::x86::UC_FP_XSTATE;
+
+      // Pointer to where the fpreg memory is
+      guest_uctx->uc_mcontext.fpregs = static_cast<uint32_t>(FPStateLocation);
+      FEXCore::x86::_libc_fpstate *fpstate = reinterpret_cast<FEXCore::x86::_libc_fpstate*>(FPStateLocation);
+
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86::FEX_REG_GS] = Frame->State.gs;
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86::FEX_REG_FS] = Frame->State.fs;
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86::FEX_REG_ES] = Frame->State.es;
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86::FEX_REG_DS] = Frame->State.ds;
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86::FEX_REG_TRAPNO] = ConvertSignalToTrapNo(Signal, HostSigInfo);
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86::FEX_REG_ERR] = ConvertSignalToError(Signal, HostSigInfo);
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86::FEX_REG_EIP] = Frame->State.rip;
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86::FEX_REG_CS] = Frame->State.cs;
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86::FEX_REG_EFL] = 0;
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86::FEX_REG_UESP] = 0;
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86::FEX_REG_SS] = Frame->State.ss;
+
+#define COPY_REG(x) \
+      guest_uctx->uc_mcontext.gregs[FEXCore::x86::FEX_REG_##x] = Frame->State.gregs[X86State::REG_##x];
+      COPY_REG(RDI);
+      COPY_REG(RSI);
+      COPY_REG(RBP);
+      COPY_REG(RBX);
+      COPY_REG(RDX);
+      COPY_REG(RAX);
+      COPY_REG(RCX);
+      COPY_REG(RSP);
+#undef COPY_REG
+
+      // Copy float registers
+      for (size_t i = 0; i < 8; ++i) {
+        // 32-bit st register size is only 10 bytes. Not padded to 16byte like x86-64
+        memcpy(&fpstate->_st[i], &Frame->State.mm[i], 10);
+      }
+
+      // Extended XMM state
+      fpstate->status = FEXCore::x86::fpstate_magic::MAGIC_XFPSTATE;
+      memcpy(fpstate->_xmm, Frame->State.xmm, sizeof(Frame->State.xmm));
+
+      // FCW store default
+      fpstate->fcw = Frame->State.FCW;
+      fpstate->ftw = Frame->State.FTW;
+      // Reconstruct FSW
+      fpstate->fsw =
+        (Frame->State.flags[FEXCore::X86State::X87FLAG_TOP_LOC] << 11) |
+        (Frame->State.flags[FEXCore::X86State::X87FLAG_C0_LOC] << 8) |
+        (Frame->State.flags[FEXCore::X86State::X87FLAG_C1_LOC] << 9) |
+        (Frame->State.flags[FEXCore::X86State::X87FLAG_C2_LOC] << 10) |
+        (Frame->State.flags[FEXCore::X86State::X87FLAG_C3_LOC] << 14);
+
+      // Copy over signal stack information
+      guest_uctx->uc_stack.ss_flags = GuestStack->ss_flags;
+      guest_uctx->uc_stack.ss_sp = static_cast<uint32_t>(reinterpret_cast<uint64_t>(GuestStack->ss_sp));
+      guest_uctx->uc_stack.ss_size = GuestStack->ss_size;
 
       // These three elements are in every siginfo
       guest_siginfo->si_signo = HostSigInfo->si_signo;
@@ -201,6 +344,12 @@ bool Dispatcher::HandleGuestSignal(int Signal, void *info, void *ucontext, Guest
       switch (Signal) {
         case SIGSEGV:
         case SIGBUS:
+          // Macro expansion to get the si_addr
+          // This is the address trying to be accessed, not the RIP
+          guest_siginfo->_sifields._sigfault.addr = static_cast<uint32_t>(reinterpret_cast<uintptr_t>(HostSigInfo->si_addr));
+          break;
+        case SIGFPE:
+        case SIGILL:
           // Macro expansion to get the si_addr
           // Can't really give a real result here. Pull from the context for now
           guest_siginfo->_sifields._sigfault.addr = Frame->State.rip;
@@ -212,40 +361,59 @@ bool Dispatcher::HandleGuestSignal(int Signal, void *info, void *ucontext, Guest
           guest_siginfo->_sifields._sigchld.utime = HostSigInfo->si_utime;
           guest_siginfo->_sifields._sigchld.stime = HostSigInfo->si_stime;
           break;
-      default:
-        LogMan::Msg::D("Unhandled siginfo_t signal: %d", Signal);
-        // Hope for the best, most things just copy over
-        memcpy(guest_siginfo, info, sizeof(siginfo_t));
-        break;
+        case SIGALRM:
+        case SIGVTALRM:
+          guest_siginfo->_sifields._timer.tid = HostSigInfo->si_timerid;
+          guest_siginfo->_sifields._timer.overrun = HostSigInfo->si_overrun;
+          guest_siginfo->_sifields._timer.sigval.sival_int = HostSigInfo->si_int;
+          break;
+        default:
+          LogMan::Msg::E("Unhandled siginfo_t for signal: %d\n", Signal);
+          break;
       }
 
       NewGuestSP -= 4;
       *(uint32_t*)NewGuestSP = UContextLocation;
       NewGuestSP -= 4;
       *(uint32_t*)NewGuestSP = SigInfoLocation;
+      NewGuestSP -= 4;
+      *(uint32_t*)NewGuestSP = Signal;
     }
 
     Frame->State.rip = reinterpret_cast<uint64_t>(GuestAction->sigaction_handler.sigaction);
   }
   else {
+    if (!Is64BitMode) {
+      NewGuestSP -= 4;
+      *(uint32_t*)NewGuestSP = Signal;
+    }
+
     Frame->State.rip = reinterpret_cast<uint64_t>(GuestAction->sigaction_handler.handler);
   }
 
-  if (CTX->Config.Is64BitMode) {
-    Frame->State.gregs[X86State::REG_RDI] = Signal;
+  if (Is64BitMode) {
+    Frame->State.gregs[FEXCore::X86State::REG_RDI] = Signal;
 
     // Set up the new SP for stack handling
     NewGuestSP -= 8;
-    *(uint64_t*)NewGuestSP = CTX->X86CodeGen.SignalReturn;
-    Frame->State.gregs[X86State::REG_RSP] = NewGuestSP;
+    *(uint64_t*)NewGuestSP = SignalReturn;
+    Frame->State.gregs[FEXCore::X86State::REG_RSP] = NewGuestSP;
   }
   else {
     NewGuestSP -= 4;
-    *(uint32_t*)NewGuestSP = CTX->X86CodeGen.SignalReturn;
-    LOGMAN_THROW_A(CTX->X86CodeGen.SignalReturn < 0x1'0000'0000ULL, "This needs to be below 4GB");
-    Frame->State.gregs[X86State::REG_RSP] = NewGuestSP;
+    *(uint32_t*)NewGuestSP = SignalReturn;
+    LOGMAN_THROW_A(SignalReturn < 0x1'0000'0000ULL, "This needs to be below 4GB");
+    Frame->State.gregs[FEXCore::X86State::REG_RSP] = NewGuestSP;
   }
 
+  // The guest starts its signal frame with a zero initialized FPU
+  // Set that up now. Little bit costly but it's a requirement
+  // This state will be restored on rt_sigreturn
+  memset(Frame->State.xmm, 0, sizeof(Frame->State.xmm));
+  memset(Frame->State.mm, 0, sizeof(Frame->State.mm));
+  Frame->State.FCW = 0x37F;
+  Frame->State.FTW = 0xFFFF;
+
   return true;
 }
 
@@ -291,9 +459,6 @@ bool Dispatcher::HandleSignalPause(int Signal, void *info, void *ucontext) {
       ArchHelpers::Context::SetPc(ucontext, ThreadPauseHandlerAddress);
     }
 
-    // Set the new PC
-    ArchHelpers::Context::SetPc(ucontext, ThreadPauseHandlerAddress);
-
     // Set our state register to point to our guest thread data
     ArchHelpers::Context::SetState(ucontext, reinterpret_cast<uint64_t>(Frame));
 

External/FEXCore/Source/Interface/Core/Dispatcher/Dispatcher.h

@@ -1,11 +1,24 @@
 #pragma once
 
 #include <FEXCore/Core/CPUBackend.h>
-#include <FEXCore/Core/SignalDelegator.h>
 
 #include "Interface/Context/Context.h"
 
+#include <bits/types/stack_t.h>
+#include <cstdint>
+#include <stddef.h>
 #include <stack>
+#include <tuple>
+#include <vector>
+
+namespace FEXCore {
+struct GuestSigAction;
+}
+
+namespace FEXCore::Core {
+struct CpuStateFrame;
+struct InternalThreadState;
+}
 
 namespace FEXCore::CPU {
 

External/FEXCore/Source/Interface/Core/Dispatcher/X86Dispatcher.cpp

@@ -1,10 +1,23 @@
+#include "Interface/Core/LookupCache.h"
+
 #include "Interface/Core/Dispatcher/X86Dispatcher.h"
 
 #include "Interface/Core/Interpreter/InterpreterClass.h"
+#include "Interface/Core/X86HelperGen.h"
 #include "Interface/Context/Context.h"
 
 #include <FEXCore/Core/X86Enums.h>
+#include <FEXCore/Core/CoreState.h>
+#include <FEXCore/Core/CPUBackend.h>
+#include <FEXCore/Debug/InternalThreadState.h>
+#include <FEXCore/Utils/Allocator.h>
+
 #include <cmath>
+#include <memory>
+#include <stddef.h>
+#include <stdint.h>
+#include <sys/mman.h>
+#include "xbyak/xbyak.h"
 
 namespace FEXCore::CPU {
 static constexpr size_t MAX_DISPATCHER_CODE_SIZE = 4096;
@@ -12,7 +25,9 @@ static constexpr size_t MAX_DISPATCHER_CODE_SIZE = 4096;
 
 X86Dispatcher::X86Dispatcher(FEXCore::Context::Context *ctx, FEXCore::Core::InternalThreadState *Thread, DispatcherConfig &config)
   : Dispatcher(ctx, Thread)
-  , Xbyak::CodeGenerator(MAX_DISPATCHER_CODE_SIZE, nullptr, this) {
+  , Xbyak::CodeGenerator(MAX_DISPATCHER_CODE_SIZE,
+      FEXCore::Allocator::mmap(nullptr, MAX_DISPATCHER_CODE_SIZE, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0),
+      nullptr) {
 
   using namespace Xbyak;
   using namespace Xbyak::util;
@@ -298,7 +313,7 @@ X86Dispatcher::X86Dispatcher(FEXCore::Context::Context *ctx, FEXCore::Core::Inte
 }
 
 X86Dispatcher::~X86Dispatcher() {
-
+  FEXCore::Allocator::munmap(top_, MAX_DISPATCHER_CODE_SIZE);
 }
 
 #ifdef _M_X86_64

External/FEXCore/Source/Interface/Core/Dispatcher/X86Dispatcher.h

@@ -2,26 +2,24 @@
 
 #include "Interface/Core/Dispatcher/Dispatcher.h"
 
-#include <FEXCore/Utils/Allocator.h>
-
 #define XBYAK64
 #include <xbyak/xbyak.h>
 
+namespace FEXCore::Context {
+  struct Context;
+}
+
+namespace FEXCore::Core {
+  struct InternalThreadState;
+}
+
 namespace FEXCore::CPU {
 
-class X86Dispatcher final : public Dispatcher, public Xbyak::CodeGenerator, public Xbyak::Allocator {
+class X86Dispatcher final : public Dispatcher, public Xbyak::CodeGenerator {
   public:
     X86Dispatcher(FEXCore::Context::Context *ctx, FEXCore::Core::InternalThreadState *Thread, DispatcherConfig &config);
 
     virtual ~X86Dispatcher() override;
-
-    // Xbyak::Allocator
-    Xbyak::uint8 *alloc(size_t size) override { Size = size; return reinterpret_cast<uint8_t*>(FEXCore::Allocator::mmap(nullptr, size, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0)); }
-    void free(Xbyak::uint8 *p) override { FEXCore::Allocator::munmap(p, Size); }
-    bool useProtect() const override { return false; }
-
-  private:
-    size_t Size{};
 };
 
 }

External/FEXCore/Source/Interface/Core/Frontend.cpp

@@ -7,17 +7,24 @@ $end_info$
 
 #include "Interface/Context/Context.h"
 #include "Interface/Core/Frontend.h"
-#include "Interface/Core/InternalThreadState.h"
 
 #include <array>
+#include <assert.h>
 #include <algorithm>
 #include <cstring>
+#include <FEXCore/Config/Config.h>
 #include <FEXCore/Core/X86Enums.h>
 #include <FEXCore/Debug/X86Tables.h>
+#include <FEXCore/HLE/SyscallHandler.h>
+#include <FEXCore/Utils/Allocator.h>
 #include <FEXCore/Utils/LogManager.h>
+#include <FEXCore/Utils/Telemetry.h>
 #include <set>
+#include <sys/mman.h>
 
 namespace FEXCore::Frontend {
+#include "Interface/Core/VSyscall/VSyscall.inc"
+
 using namespace FEXCore::X86Tables;
 
 static uint32_t MapModRMToReg(uint8_t REX, uint8_t bits, bool HighBits, bool HasREX, bool HasXMM, bool HasMM, uint8_t InvalidOffset = 16) {
@@ -121,8 +128,18 @@ static uint32_t MapModRMToReg(uint8_t REX, uint8_t bits, bool HighBits, bool Has
 }
 
 Decoder::Decoder(FEXCore::Context::Context *ctx)
-  : CTX {ctx} {
-  DecodedBuffer.resize(DefaultDecodedBufferSize);
+  : CTX {ctx}
+  , OSABI { ctx->SyscallHandler ? ctx->SyscallHandler->GetOSABI() : FEXCore::HLE::SyscallOSABI::OS_UNKNOWN } {
+  // Using mmap is a start-up time optimization
+  // Take advantage of page faulting to reduce startup time for minimal runtime cost
+  DecodedBuffer =
+    reinterpret_cast<FEXCore::X86Tables::DecodedInst *>(
+      FEXCore::Allocator::mmap(0, sizeof(FEXCore::X86Tables::DecodedInst) * DefaultDecodedBufferSize,
+      PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0));
+}
+
+Decoder::~Decoder() {
+  FEXCore::Allocator::munmap(DecodedBuffer, sizeof(FEXCore::X86Tables::DecodedInst) * DefaultDecodedBufferSize);
 }
 
 uint8_t Decoder::ReadByte() {
@@ -214,28 +231,28 @@ void Decoder::DecodeModRM_16(X86Tables::DecodedOperand *Operand, X86Tables::ModR
     {FEXCore::X86State::REG_RBX, FEXCore::X86State::REG_RDI},
     {FEXCore::X86State::REG_RBP, FEXCore::X86State::REG_RSI},
     {FEXCore::X86State::REG_RBP, FEXCore::X86State::REG_RDI},
-    {FEXCore::X86State::REG_RSI, 255},
-    {FEXCore::X86State::REG_RDI, 255},
-    {255, 255},
-    {FEXCore::X86State::REG_RBX, 255},
+    {FEXCore::X86State::REG_RSI, FEXCore::X86State::REG_INVALID},
+    {FEXCore::X86State::REG_RDI, FEXCore::X86State::REG_INVALID},
+    {FEXCore::X86State::REG_INVALID, FEXCore::X86State::REG_INVALID},
+    {FEXCore::X86State::REG_RBX, FEXCore::X86State::REG_INVALID},
     // Mod = 0b01
     {FEXCore::X86State::REG_RBX, FEXCore::X86State::REG_RSI},
     {FEXCore::X86State::REG_RBX, FEXCore::X86State::REG_RDI},
     {FEXCore::X86State::REG_RBP, FEXCore::X86State::REG_RSI},
     {FEXCore::X86State::REG_RBP, FEXCore::X86State::REG_RDI},
-    {FEXCore::X86State::REG_RSI, 255},
-    {FEXCore::X86State::REG_RDI, 255},
-    {FEXCore::X86State::REG_RBP, 255},
-    {FEXCore::X86State::REG_RBX, 255},
+    {FEXCore::X86State::REG_RSI, FEXCore::X86State::REG_INVALID},
+    {FEXCore::X86State::REG_RDI, FEXCore::X86State::REG_INVALID},
+    {FEXCore::X86State::REG_RBP, FEXCore::X86State::REG_INVALID},
+    {FEXCore::X86State::REG_RBX, FEXCore::X86State::REG_INVALID},
     // Mod = 0b10
     {FEXCore::X86State::REG_RBX, FEXCore::X86State::REG_RSI},
     {FEXCore::X86State::REG_RBX, FEXCore::X86State::REG_RDI},
     {FEXCore::X86State::REG_RBP, FEXCore::X86State::REG_RSI},
     {FEXCore::X86State::REG_RBP, FEXCore::X86State::REG_RDI},
-    {FEXCore::X86State::REG_RSI, 255},
-    {FEXCore::X86State::REG_RDI, 255},
-    {FEXCore::X86State::REG_RBP, 255},
-    {FEXCore::X86State::REG_RBX, 255},
+    {FEXCore::X86State::REG_RSI, FEXCore::X86State::REG_INVALID},
+    {FEXCore::X86State::REG_RDI, FEXCore::X86State::REG_INVALID},
+    {FEXCore::X86State::REG_RBP, FEXCore::X86State::REG_INVALID},
+    {FEXCore::X86State::REG_RBX, FEXCore::X86State::REG_INVALID},
   }};
 
   uint8_t LookupIndex = ModRM.mod << 3 | ModRM.rm;
@@ -683,6 +700,7 @@ bool Decoder::NormalOpHeader(FEXCore::X86Tables::X86InstInfo const *Info, uint16
     return NormalOp(&X87Ops[X87Op], X87Op);
   }
   else if (Info->Type == FEXCore::X86Tables::TYPE_VEX_TABLE_PREFIX) {
+    FEXCORE_TELEMETRY_SET(VEXOpTelem, 1);
     uint16_t map_select = 1;
     uint16_t pp = 0;
 
@@ -710,6 +728,7 @@ bool Decoder::NormalOpHeader(FEXCore::X86Tables::X86InstInfo const *Info, uint16
 
     if (LocalInfo->Type >= FEXCore::X86Tables::TYPE_VEX_GROUP_12 &&
         LocalInfo->Type <= FEXCore::X86Tables::TYPE_VEX_GROUP_17) {
+    FEXCORE_TELEMETRY_SET(VEXOpTelem, 1);
     // We have ModRM
     uint8_t ModRMByte = ReadByte();
     DecodeInst->ModRM = ModRMByte;
@@ -727,6 +746,8 @@ bool Decoder::NormalOpHeader(FEXCore::X86Tables::X86InstInfo const *Info, uint16
       return NormalOp(LocalInfo, Op);
   }
   else if (Info->Type == FEXCore::X86Tables::TYPE_GROUP_EVEX) {
+    FEXCORE_TELEMETRY_SET(EVEXOpTelem, 1);
+
     /* uint8_t P1 = */ ReadByte();
     /* uint8_t P2 = */ ReadByte();
     /* uint8_t P3 = */ ReadByte();
@@ -929,7 +950,7 @@ bool Decoder::DecodeInstruction(uint64_t PC) {
   }
 
   if (DecodeInst->Dest.IsGPR()) {
-    assert(DecodeInst->Dest.Data.GPR.GPR != 255);
+    assert(DecodeInst->Dest.Data.GPR.GPR != FEXCore::X86State::REG_INVALID);
   }
 
   return true;
@@ -1004,6 +1025,25 @@ void Decoder::BranchTargetInMultiblockRange() {
   }
 }
 
+const uint8_t *Decoder::AdjustAddrForSpecialRegion(uint8_t const* _InstStream, uint64_t EntryPoint, uint64_t RIP) {
+  constexpr uint64_t VSyscall_Base = 0xFFFF'FFFF'FF60'0000ULL;
+  constexpr uint64_t VSyscall_End = VSyscall_Base + 0x1000;
+
+  if (OSABI == FEXCore::HLE::SyscallOSABI::OS_LINUX64 &&
+      RIP >= VSyscall_Base &&
+      RIP < VSyscall_End) {
+    // VSyscall
+    // This doesn't exist on AArch64 and on x86_64 hosts this is emulated with faults to a region mapped with --xp permissions
+    // Offset     0: vgettimeofday
+    // Offset 0x400: vtime
+    // Offset 0x800: vgetcpu
+    uint64_t Offset = RIP - VSyscall_Base;
+    return VSyscallData + Offset;
+  }
+
+  return _InstStream - EntryPoint + RIP;
+}
+
 bool Decoder::DecodeInstructionsAtEntry(uint8_t const* _InstStream, uint64_t PC) {
   Blocks.clear();
   BlocksToDecode.clear();
@@ -1047,7 +1087,7 @@ bool Decoder::DecodeInstructionsAtEntry(uint8_t const* _InstStream, uint64_t PC)
     uint64_t BlockStartOffset = DecodedSize;
 
     // Do a bit of pointer math to figure out where we are in code
-    InstStream = _InstStream - EntryPoint + RIPToDecode;
+    InstStream = AdjustAddrForSpecialRegion(_InstStream, EntryPoint, RIPToDecode);
 
     while (1) {
       ErrorDuringDecoding = !DecodeInstruction(RIPToDecode + PCOffset);
@@ -1091,7 +1131,7 @@ bool Decoder::DecodeInstructionsAtEntry(uint8_t const* _InstStream, uint64_t PC)
       }
 
       if (DecodedSize >= CTX->Config.MaxInstPerBlock ||
-          DecodedSize >= DecodedBuffer.size()) {
+          DecodedSize >= DefaultDecodedBufferSize) {
         break;
       }
 
@@ -1108,7 +1148,7 @@ bool Decoder::DecodeInstructionsAtEntry(uint8_t const* _InstStream, uint64_t PC)
 
     // Copy over only the number of instructions we decoded
     CurrentBlockDecoding.NumInstructions = BlockNumberOfInstructions;
-    CurrentBlockDecoding.DecodedInstructions = &DecodedBuffer.at(BlockStartOffset);
+    CurrentBlockDecoding.DecodedInstructions = &DecodedBuffer[BlockStartOffset];
   }
 
 

External/FEXCore/Source/Interface/Core/Frontend.h

@@ -1,11 +1,13 @@
 #pragma once
 
 #include <FEXCore/Debug/X86Tables.h>
+#include <FEXCore/HLE/SyscallHandler.h>
+#include <FEXCore/Utils/Telemetry.h>
+
 #include <array>
 #include <cstdint>
-#include <utility>
 #include <set>
-#include <stack>
+#include <stddef.h>
 #include <vector>
 
 namespace FEXCore::Context {
@@ -24,6 +26,7 @@ public:
   };
 
   Decoder(FEXCore::Context::Context *ctx);
+  ~Decoder();
   bool DecodeInstructionsAtEntry(uint8_t const* InstStream, uint64_t PC);
 
   std::vector<DecodedBlocks> const *GetDecodedBlocks() const {
@@ -37,6 +40,7 @@ public:
   void SetExternalBranches(std::set<uint64_t> *v) { ExternalBranches = v; }
 private:
   FEXCore::Context::Context *CTX;
+  const FEXCore::HLE::SyscallOSABI OSABI{};
 
   bool DecodeInstruction(uint64_t PC);
 
@@ -50,7 +54,7 @@ private:
   bool NormalOpHeader(FEXCore::X86Tables::X86InstInfo const *Info, uint16_t Op);
 
   static constexpr size_t DefaultDecodedBufferSize = 0x10000;
-  std::vector<FEXCore::X86Tables::DecodedInst> DecodedBuffer;
+  FEXCore::X86Tables::DecodedInst *DecodedBuffer{};
   size_t DecodedSize {};
 
   uint8_t const *InstStream;
@@ -83,5 +87,10 @@ private:
     &FEXCore::Frontend::Decoder::DecodeModRM_64,
     &FEXCore::Frontend::Decoder::DecodeModRM_16,
   };
+
+  const uint8_t *AdjustAddrForSpecialRegion(uint8_t const* _InstStream, uint64_t EntryPoint, uint64_t RIP);
+
+  FEXCORE_TELEMETRY_INIT(VEXOpTelem, TYPE_USES_VEX_OPS);
+  FEXCORE_TELEMETRY_INIT(EVEXOpTelem, TYPE_USES_EVEX_OPS);
 };
 }

External/FEXCore/Source/Interface/Core/GdbServer.cpp

@@ -8,30 +8,41 @@ $end_info$
 #include <cstdlib>
 #include <cstdio>
 #include <iomanip>
-#include <iostream>
 #include <sstream>
 #include <string>
 #include <memory>
 #include <optional>
 #include "Common/NetStream.h"
 #include "Common/SoftFloat.h"
+#include "Interface/Context/Context.h"
 
+#include <FEXCore/Config/Config.h>
+#include <FEXCore/Core/Context.h>
+#include <FEXCore/Core/CoreState.h>
+#include <FEXCore/Core/SignalDelegator.h>
+#include <FEXCore/Core/X86Enums.h>
+#include <FEXCore/Debug/InternalThreadState.h>
+#include <FEXCore/HLE/Linux/ThreadManagement.h>
 #include <FEXCore/Utils/CompilerDefs.h>
 #include <FEXCore/Utils/LogManager.h>
+#include <FEXCore/Utils/Threads.h>
 
+#include <atomic>
 #include <cstring>
+#include <errno.h>
 #include <fcntl.h>
-#include <fmt/format.h>
 #include <fstream>
+#include <fmt/format.h>
 #include <netdb.h>
+#include <signal.h>
+#include <stddef.h>
+#include <string_view>
 #include <sys/socket.h>
-#include <sys/types.h>
 #include <unistd.h>
-
+#include <utility>
+#include <vector>
 
 #include "GdbServer.h"
-#include <FEXCore/Core/CodeLoader.h>
-#include <FEXCore/Core/X86Enums.h>
 
 namespace FEXCore
 {
@@ -939,7 +950,7 @@ void GdbServer::GdbServerLoop() {
 
     {
         std::lock_guard lk(sendMutex);
-        CommsStream.release();
+        CommsStream.reset();
     }
   }
 }

External/FEXCore/Source/Interface/Core/GdbServer.h

@@ -5,18 +5,21 @@ $end_info$
 */
 #pragma once
 
-#include <mutex>
-#include <thread>
-
-#include "Interface/Context/Context.h"
-#include "Common/NetStream.h"
-
+#include <FEXCore/Config/Config.h>
 #include <FEXCore/Utils/Threads.h>
 
+#include <istream>
+#include <memory>
 #include <mutex>
+#include <stdint.h>
+#include <string>
 
 namespace FEXCore {
 
+namespace Context {
+  struct Context;
+}
+
 class GdbServer {
 public:
     GdbServer(FEXCore::Context::Context *ctx);

External/FEXCore/Source/Interface/Core/Interpreter/InterpreterClass.h

@@ -1,6 +1,5 @@
 #pragma once
 
-#include "Interface/Core/LookupCache.h"
 #include "Interface/Core/InternalThreadState.h"
 #include "Interface/Core/Dispatcher/Dispatcher.h"
 

External/FEXCore/Source/Interface/Core/Interpreter/InterpreterCore.cpp

@@ -1,30 +1,31 @@
-#include "Common/MathUtils.h"
-#include "Common/SoftFloat.h"
 #include "Interface/Context/Context.h"
 
 #include "Interface/Core/ArchHelpers/Arm64.h"
 #include "Interface/Core/ArchHelpers/MContext.h"
-#include "Interface/Core/LookupCache.h"
-#include "Interface/Core/DebugData.h"
-#include "Interface/Core/InternalThreadState.h"
+#include "Interface/Core/Dispatcher/Dispatcher.h"
 #include "Interface/Core/Interpreter/InterpreterClass.h"
+#include <FEXCore/Config/Config.h>
+#include <FEXCore/Core/CoreState.h>
+#include <FEXCore/Core/SignalDelegator.h>
+#include <FEXCore/Debug/InternalThreadState.h>
 #include <FEXCore/Utils/LogManager.h>
 
-#include <FEXCore/Core/CPUBackend.h>
-#include <FEXCore/HLE/SyscallHandler.h>
-#include <FEXCore/IR/IR.h>
-#include <FEXCore/IR/IntrusiveIRList.h>
-
-#include "Interface/HLE/Thunks/Thunks.h"
-
-#include <atomic>
-#include <cmath>
-#include <limits>
-#include <vector>
+#include <memory>
+#include <bits/types/stack_t.h>
+#include <signal.h>
+#include <stdint.h>
+#include <unordered_map>
+#include <utility>
 
 #include "InterpreterOps.h"
 
+namespace FEXCore::IR {
+  class IRListView;
+  class RegisterAllocationData;
+}
+
 namespace FEXCore::CPU {
+class CPUBackend;
 
 static void InterpreterExecution(FEXCore::Core::CpuStateFrame *Frame) {
   auto Thread = Frame->Thread;
@@ -78,6 +79,18 @@ bool InterpreterCore::HandleSIGBUS(int Signal, void *info, void *ucontext) {
         return false;
       }
     }
+    else if ((Instr & FEXCore::ArchHelpers::Arm64::LDAXR_MASK) == FEXCore::ArchHelpers::Arm64::LDAXR_INST) { // LDAXR*
+      uint64_t BytesToSkip = FEXCore::ArchHelpers::Arm64::HandleAtomicLoadstoreExclusive(ucontext, info);
+      if (BytesToSkip) {
+        // Skip this instruction now
+        ArchHelpers::Context::SetPc(ucontext, ArchHelpers::Context::GetPc(ucontext) + BytesToSkip);
+        return true;
+      }
+      else {
+        LogMan::Msg::E("Unhandled JIT SIGBUS LDAXR: PC: %p Instruction: 0x%08x\n", PC, PC[0]);
+        return false;
+      }
+    }
   }
   return false;
 }
@@ -105,7 +118,7 @@ InterpreterCore::InterpreterCore(FEXCore::Context::Context *ctx, FEXCore::Core::
       return Core->Dispatcher->HandleGuestSignal(Signal, info, ucontext, GuestAction, GuestStack);
     };
 
-    for (uint32_t Signal = 0; Signal < SignalDelegator::MAX_SIGNALS; ++Signal) {
+    for (uint32_t Signal = 0; Signal <= SignalDelegator::MAX_SIGNALS; ++Signal) {
       CTX->SignalDelegation->RegisterHostSignalHandlerForGuest(Signal, GuestSignalHandler);
     }
   }

External/FEXCore/Source/Interface/Core/Interpreter/InterpreterOps.h

@@ -1,9 +1,13 @@
+#pragma once
+#include <stdint.h>
+
 namespace FEXCore::Core {
   struct InternalThreadState;
 }
 
 namespace FEXCore::IR {
     class IRListView;
+    struct IROp_Header;
 }
 
 namespace FEXCore::Core{
@@ -32,11 +36,11 @@ namespace FEXCore::CPU {
     FallbackABI ABI;
     void *fn;
   };
-  
+
   class InterpreterOps {
 
     public:
       static void InterpretIR(FEXCore::Core::InternalThreadState *Thread, uint64_t Entry, FEXCore::IR::IRListView *CurrentIR, FEXCore::Core::DebugData *DebugData);
       static bool GetFallbackHandler(IR::IROp_Header *IROp, FallbackInfo *Info);
   };
-};
+};

External/FEXCore/Source/Interface/Core/JIT/Arm64/ALUOps.cpp

@@ -46,7 +46,7 @@ DEF_OP(TruncElementPair) {
       mov(Dst.second, Src.second);
       break;
     }
-    default: LOGMAN_MSG_A("Unhandled Truncation size: %d", Op->Size); break;
+    default: LOGMAN_MSG_A_FMT("Unhandled Truncation size: {}", Op->Size); break;
   }
 }
 
@@ -95,7 +95,7 @@ DEF_OP(Add) {
       case 8:
         add(GetReg<RA_64>(Node), GetReg<RA_64>(Op->Header.Args[0].ID()), Const);
         break;
-      default: LOGMAN_MSG_A("Unsupported Add size: %d", OpSize);
+      default: LOGMAN_MSG_A_FMT("Unsupported Add size: {}", OpSize);
     }
   } else {
     switch (OpSize) {
@@ -105,7 +105,7 @@ DEF_OP(Add) {
       case 8:
         add(GetReg<RA_64>(Node), GetReg<RA_64>(Op->Header.Args[0].ID()), GetReg<RA_64>(Op->Header.Args[1].ID()));
         break;
-      default: LOGMAN_MSG_A("Unsupported Add size: %d", OpSize);
+      default: LOGMAN_MSG_A_FMT("Unsupported Add size: {}", OpSize);
     }
   }
 }
@@ -121,7 +121,7 @@ DEF_OP(Sub) {
       case 8:
         sub(GRS(Node), GRS(Op->Header.Args[0].ID()), Const);
         break;
-      default: LOGMAN_MSG_A("Unsupported Sub size: %d", OpSize);
+      default: LOGMAN_MSG_A_FMT("Unsupported Sub size: {}", OpSize);
     }
   } else {
     switch (OpSize) {
@@ -131,7 +131,7 @@ DEF_OP(Sub) {
       case 8:
         sub(GetReg<RA_64>(Node), GetReg<RA_64>(Op->Header.Args[0].ID()), GetReg<RA_64>(Op->Header.Args[1].ID()));
         break;
-      default: LOGMAN_MSG_A("Unsupported Sub size: %d", OpSize);
+      default: LOGMAN_MSG_A_FMT("Unsupported Sub size: {}", OpSize);
     }
   }
 
@@ -147,7 +147,7 @@ DEF_OP(Neg) {
     case 8:
       neg(GetReg<RA_64>(Node), GetReg<RA_64>(Op->Header.Args[0].ID()));
       break;
-    default: LOGMAN_MSG_A("Unsupported Not size: %d", OpSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported Neg size: {}", OpSize);
   }
 }
 
@@ -163,7 +163,7 @@ DEF_OP(Mul) {
     case 8:
       mul(Dst, GetReg<RA_64>(Op->Header.Args[0].ID()), GetReg<RA_64>(Op->Header.Args[1].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown Sext size: %d", OpSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Mul size: %d", OpSize);
   }
 }
 
@@ -179,7 +179,7 @@ DEF_OP(UMul) {
     case 8:
       mul(Dst, GetReg<RA_64>(Op->Header.Args[0].ID()), GetReg<RA_64>(Op->Header.Args[1].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown Sext size: %d", OpSize);
+    default: LOGMAN_MSG_A_FMT("Unknown UMul size: {}", OpSize);
   }
 }
 
@@ -216,7 +216,7 @@ DEF_OP(Div) {
       sdiv(GetReg<RA_64>(Node), GetReg<RA_64>(Op->Header.Args[0].ID()), GetReg<RA_64>(Op->Header.Args[1].ID()));
     break;
     }
-    default: LOGMAN_MSG_A("Unknown DIV Size: %d", Size); break;
+    default: LOGMAN_MSG_A_FMT("Unknown DIV Size: {}", Size); break;
   }
 }
 
@@ -243,7 +243,7 @@ DEF_OP(UDiv) {
       udiv(GetReg<RA_64>(Node), GetReg<RA_64>(Op->Header.Args[0].ID()), GetReg<RA_64>(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown UDIV Size: %d", Size); break;
+    default: LOGMAN_MSG_A_FMT("Unknown UDIV Size: {}", Size); break;
   }
 }
 
@@ -290,7 +290,7 @@ DEF_OP(Rem) {
       msub(GetReg<RA_64>(Node), TMP1, Divisor, Dividend);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown REM Size: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown REM Size: {}", OpSize); break;
   }
 }
 
@@ -332,7 +332,7 @@ DEF_OP(URem) {
       msub(GetReg<RA_64>(Node), TMP1, Divisor, Dividend);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown UREM Size: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown UREM Size: {}", OpSize); break;
   }
 }
 
@@ -349,7 +349,7 @@ DEF_OP(MulH) {
     case 8:
       smulh(GetReg<RA_64>(Node), GetReg<RA_64>(Op->Header.Args[0].ID()), GetReg<RA_64>(Op->Header.Args[1].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown Sext size: %d", OpSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Sext size: {}", OpSize);
   }
 }
 
@@ -366,7 +366,7 @@ DEF_OP(UMulH) {
     case 8:
       umulh(GetReg<RA_64>(Node), GetReg<RA_64>(Op->Header.Args[0].ID()), GetReg<RA_64>(Op->Header.Args[1].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown Sext size: %d", OpSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Sext size: {}", OpSize);
   }
 }
 
@@ -462,7 +462,7 @@ DEF_OP(Ror) {
       break;
       }
 
-      default: LOGMAN_MSG_A("Unhandled ROR size: %d", OpSize);
+      default: LOGMAN_MSG_A_FMT("Unhandled ROR size: {}", OpSize);
     }
   } else {
     switch (OpSize) {
@@ -475,7 +475,7 @@ DEF_OP(Ror) {
       break;
       }
 
-      default: LOGMAN_MSG_A("Unhandled ROR size: %d", OpSize);
+      default: LOGMAN_MSG_A_FMT("Unhandled ROR size: {}", OpSize);
     }
   }
 }
@@ -494,7 +494,7 @@ DEF_OP(Extr) {
     break;
     }
 
-    default: LOGMAN_MSG_A("Unhandled EXTR size: %d", OpSize);
+    default: LOGMAN_MSG_A_FMT("Unhandled EXTR size: {}", OpSize);
   }
 }
 
@@ -539,7 +539,7 @@ DEF_OP(LDiv) {
       mov(GetReg<RA_64>(Node), x0);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown LDIV Size: %d", Size); break;
+    default: LOGMAN_MSG_A_FMT("Unknown LDIV Size: {}", Size); break;
   }
 }
 
@@ -582,7 +582,7 @@ DEF_OP(LUDiv) {
       mov(GetReg<RA_64>(Node), x0);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown LUDIV Size: %d", Size); break;
+    default: LOGMAN_MSG_A_FMT("Unknown LUDIV Size: {}", Size); break;
   }
 }
 
@@ -635,7 +635,7 @@ DEF_OP(LRem) {
       mov(GetReg<RA_64>(Node), x0);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown LREM Size: %d", Size); break;
+    default: LOGMAN_MSG_A_FMT("Unknown LREM Size: {}", Size); break;
   }
 }
 
@@ -685,7 +685,7 @@ DEF_OP(LURem) {
       mov(GetReg<RA_64>(Node), x0);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown LUREM Size: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown LUREM Size: {}", OpSize); break;
   }
 }
 
@@ -699,7 +699,7 @@ DEF_OP(Not) {
     case 8:
       mvn(GetReg<RA_64>(Node), GetReg<RA_64>(Op->Header.Args[0].ID()));
       break;
-    default: LOGMAN_MSG_A("Unsupported Not size: %d", OpSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported Not size: {}", OpSize);
   }
 }
 
@@ -730,7 +730,7 @@ DEF_OP(Popcount) {
       // fmov has zero extended, unused bytes are zero
       addv(VTMP1.B(), VTMP1.V8B());
       break;
-    default: LOGMAN_MSG_A("Unsupported Popcount size: %d", OpSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported Popcount size: {}", OpSize);
   }
 
   auto Dst = GetReg<RA_32>(Node);
@@ -779,7 +779,7 @@ DEF_OP(FindMSB) {
       clz(Dst, GetReg<RA_64>(Op->Header.Args[0].ID()));
       sub(Dst, TMP1, Dst);
       break;
-    default: LOGMAN_MSG_A("Unknown REV size: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown FindMSB size: {}", OpSize); break;
   }
 }
 
@@ -800,7 +800,7 @@ DEF_OP(FindTrailingZeros) {
       rbit(GetReg<RA_64>(Node), GetReg<RA_64>(Op->Header.Args[0].ID()));
       clz(GetReg<RA_64>(Node), GetReg<RA_64>(Node));
       break;
-    default: LOGMAN_MSG_A("Unknown size: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown FindTrailingZeros size: {}", OpSize); break;
   }
 }
 
@@ -819,7 +819,7 @@ DEF_OP(CountLeadingZeroes) {
     case 8:
       clz(GetReg<RA_64>(Node), GetReg<RA_64>(Op->Header.Args[0].ID()));
       break;
-    default: LOGMAN_MSG_A("Unknown size: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown CountLeadingZeroes size: {}", OpSize); break;
   }
 }
 
@@ -837,7 +837,7 @@ DEF_OP(Rev) {
     case 8:
       rev(GetReg<RA_64>(Node), GetReg<RA_64>(Op->Header.Args[0].ID()));
       break;
-    default: LOGMAN_MSG_A("Unknown REV size: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown REV size: {}", OpSize); break;
   }
 }
 
@@ -859,14 +859,14 @@ DEF_OP(Bfi) {
       bfi(TMP1, GetReg<RA_64>(Op->Header.Args[1].ID()), Op->lsb, Op->Width);
       mov(GetReg<RA_64>(Node), TMP1);
       break;
-    default: LOGMAN_MSG_A("Unknown BFI size: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown BFI size: {}", OpSize); break;
   }
 }
 
 DEF_OP(Bfe) {
   auto Op = IROp->C<IR::IROp_Bfe>();
-  LOGMAN_THROW_A(IROp->Size <= 8, "OpSize is too large for BFE: %d", IROp->Size);
-  LOGMAN_THROW_A(Op->Width != 0, "Invalid BFE width of 0");
+  LOGMAN_THROW_A_FMT(IROp->Size <= 8, "OpSize is too large for BFE: {}", IROp->Size);
+  LOGMAN_THROW_A_FMT(Op->Width != 0, "Invalid BFE width of 0");
 
   auto Dst = GetReg<RA_64>(Node);
   ubfx(Dst, GetReg<RA_64>(Op->Header.Args[0].ID()), Op->lsb, Op->Width);
@@ -880,7 +880,7 @@ DEF_OP(Sbfe) {
   if (OpSize == 8) {
     sbfx(Dst, GetReg<RA_64>(Op->Header.Args[0].ID()), Op->lsb, Op->Width);
   } else {
-    LogMan::Msg::D("Unimplemented Sbfe size");
+    LogMan::Msg::DFmt("Unimplemented Sbfe size");
   }
 }
 
@@ -911,7 +911,7 @@ Condition MapSelectCC(IR::CondClassType Cond) {
   case FEXCore::IR::COND_MI:
   case FEXCore::IR::COND_PL:
   default:
-  LOGMAN_MSG_A("Unsupported compare type");
+  LOGMAN_MSG_A_FMT("Unsupported compare type");
   return Condition::nv;
   }
 }
@@ -929,7 +929,7 @@ DEF_OP(Select) {
   } else if (IsFPR(Op->Cmp1.ID())) {
     fcmp(GRFCMP(Op->Cmp1.ID()), GRFCMP(Op->Cmp2.ID()));
   } else {
-    LOGMAN_MSG_A("Select: Expected GPR or FPR");
+    LOGMAN_MSG_A_FMT("Select: Expected GPR or FPR");
   }
 
   auto cc = MapSelectCC(Op->Cond);
@@ -940,7 +940,7 @@ DEF_OP(Select) {
 
   if (is_const_true || is_const_false) {
     if (is_const_false != true || is_const_true != true || const_true != 1 || const_false != 0) {
-      LOGMAN_MSG_A("Select: Unsupported compare inline parameters");
+      LOGMAN_MSG_A_FMT("Select: Unsupported compare inline parameters");
     }
     cset(GRS(Node), cc);
   } else {
@@ -964,7 +964,7 @@ DEF_OP(VExtractToGPR) {
     case 8:
       umov(GetReg<RA_64>(Node), GetSrc(Op->Header.Args[0].ID()).V2D(), Op->Idx);
     break;
-    default:  LOGMAN_MSG_A("Unhandled ExtractElementSize: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled ExtractElementSize: {}", OpSize);
   }
 }
 
@@ -1027,7 +1027,7 @@ DEF_OP(FCmp) {
   bool set = false;
 
   if (Op->Flags & (1 << IR::FCMP_FLAG_EQ)) {
-    LOGMAN_THROW_A(IR::FCMP_FLAG_EQ == 0, "IR::FCMP_FLAG_EQ must equal 0");
+    LOGMAN_THROW_A_FMT(IR::FCMP_FLAG_EQ == 0, "IR::FCMP_FLAG_EQ must equal 0");
     // EQ or unordered
     cset(Dst, Condition::eq); // Z = 1
     csinc(Dst, Dst, xzr, Condition::vc); // IF !V ? Z : 1

External/FEXCore/Source/Interface/Core/JIT/Arm64/AtomicOps.cpp

@@ -34,7 +34,7 @@ DEF_OP(CASPair) {
       mov(Dst.first, TMP3);
       mov(Dst.second, TMP4);
       break;
-    default: LOGMAN_MSG_A("Unsupported: %d", OpSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported: {}", OpSize);
     }
   }
   else {
@@ -44,6 +44,7 @@ DEF_OP(CASPair) {
         aarch64::Label LoopNotExpected;
         aarch64::Label LoopExpected;
         bind(&LoopTop);
+
         ldaxp(TMP2.W(), TMP3.W(), MemOperand(MemSrc));
         cmp(TMP2.W(), Expected.first.W());
         ccmp(TMP3.W(), Expected.second.W(), NoFlag, Condition::eq);
@@ -69,6 +70,7 @@ DEF_OP(CASPair) {
         aarch64::Label LoopNotExpected;
         aarch64::Label LoopExpected;
         bind(&LoopTop);
+
         ldaxp(TMP2.X(), TMP3.X(), MemOperand(MemSrc));
         cmp(TMP2.X(), Expected.first.X());
         ccmp(TMP3.X(), Expected.second.X(), NoFlag, Condition::eq);
@@ -89,7 +91,7 @@ DEF_OP(CASPair) {
         bind(&LoopExpected);
         break;
       }
-      default: LOGMAN_MSG_A("Unsupported: %d", OpSize);
+      default: LOGMAN_MSG_A_FMT("Unsupported: {}", OpSize);
     }
   }
 }
@@ -115,7 +117,7 @@ DEF_OP(CAS) {
     case 2: casalh(TMP2.W(), Desired.W(), MemOperand(MemSrc)); break;
     case 4: casal(TMP2.W(), Desired.W(), MemOperand(MemSrc)); break;
     case 8: casal(TMP2.X(), Desired.X(), MemOperand(MemSrc)); break;
-    default: LOGMAN_MSG_A("Unsupported: %d", OpSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported: {}", OpSize);
     }
     mov(GetReg<RA_64>(Node), TMP2);
   }
@@ -206,7 +208,7 @@ DEF_OP(CAS) {
 
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", OpSize);
+      default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", OpSize);
     }
   }
 }
@@ -222,7 +224,7 @@ DEF_OP(AtomicAdd) {
     case 2: staddlh(GetReg<RA_32>(Op->Header.Args[1].ID()), MemOperand(MemSrc)); break;
     case 4: staddl(GetReg<RA_32>(Op->Header.Args[1].ID()), MemOperand(MemSrc)); break;
     case 8: staddl(GetReg<RA_64>(Op->Header.Args[1].ID()), MemOperand(MemSrc)); break;
-    default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
   else {
@@ -264,7 +266,7 @@ DEF_OP(AtomicAdd) {
         cbnz(TMP2, &LoopTop);
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
 }
@@ -281,7 +283,7 @@ DEF_OP(AtomicSub) {
     case 2: staddlh(TMP2.W(), MemOperand(MemSrc)); break;
     case 4: staddl(TMP2.W(), MemOperand(MemSrc)); break;
     case 8: staddl(TMP2.X(), MemOperand(MemSrc)); break;
-    default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
   else {
@@ -323,7 +325,7 @@ DEF_OP(AtomicSub) {
         cbnz(TMP2, &LoopTop);
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
 }
@@ -340,7 +342,7 @@ DEF_OP(AtomicAnd) {
     case 2: stclrlh(TMP2.W(), MemOperand(MemSrc)); break;
     case 4: stclrl(TMP2.W(), MemOperand(MemSrc)); break;
     case 8: stclrl(TMP2.X(), MemOperand(MemSrc)); break;
-    default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
   else {
@@ -382,7 +384,7 @@ DEF_OP(AtomicAnd) {
         cbnz(TMP2, &LoopTop);
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
 }
@@ -398,7 +400,7 @@ DEF_OP(AtomicOr) {
     case 2: stsetlh(GetReg<RA_32>(Op->Header.Args[1].ID()), MemOperand(MemSrc)); break;
     case 4: stsetl(GetReg<RA_32>(Op->Header.Args[1].ID()), MemOperand(MemSrc)); break;
     case 8: stsetl(GetReg<RA_64>(Op->Header.Args[1].ID()), MemOperand(MemSrc)); break;
-    default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
   else {
@@ -440,7 +442,7 @@ DEF_OP(AtomicOr) {
         cbnz(TMP2, &LoopTop);
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
 }
@@ -456,7 +458,7 @@ DEF_OP(AtomicXor) {
     case 2: steorlh(GetReg<RA_32>(Op->Header.Args[1].ID()), MemOperand(MemSrc)); break;
     case 4: steorl(GetReg<RA_32>(Op->Header.Args[1].ID()), MemOperand(MemSrc)); break;
     case 8: steorl(GetReg<RA_64>(Op->Header.Args[1].ID()), MemOperand(MemSrc)); break;
-    default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
   else {
@@ -498,7 +500,7 @@ DEF_OP(AtomicXor) {
         cbnz(TMP2, &LoopTop);
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
 }
@@ -515,36 +517,35 @@ DEF_OP(AtomicSwap) {
     case 2: swplh(TMP2.W(), GetReg<RA_32>(Node), MemOperand(MemSrc)); break;
     case 4: swpl(TMP2.W(), GetReg<RA_32>(Node), MemOperand(MemSrc)); break;
     case 8: swpl(TMP2.X(), GetReg<RA_64>(Node), MemOperand(MemSrc)); break;
-    default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
   else {
     // TMP2-TMP3
-    mov(TMP3, GetReg<RA_64>(Op->Header.Args[1].ID()));
     switch (Op->Size) {
       case 1: {
         aarch64::Label LoopTop;
         bind(&LoopTop);
         ldaxrb(TMP2.W(), MemOperand(MemSrc));
-        stlxrb(TMP4.W(), TMP3.W(), MemOperand(MemSrc));
+        stlxrb(TMP4.W(), GetReg<RA_32>(Op->Header.Args[1].ID()), MemOperand(MemSrc));
         cbnz(TMP4.W(), &LoopTop);
-        uxtb(GetReg<RA_64>(Node), TMP2.W());
+        uxtb(GetReg<RA_32>(Node), TMP2.W());
         break;
       }
       case 2: {
         aarch64::Label LoopTop;
         bind(&LoopTop);
         ldaxrh(TMP2.W(), MemOperand(MemSrc));
-        stlxrh(TMP4.W(), TMP3.W(), MemOperand(MemSrc));
+        stlxrh(TMP4.W(), GetReg<RA_32>(Op->Header.Args[1].ID()), MemOperand(MemSrc));
         cbnz(TMP4.W(), &LoopTop);
-        uxtw(GetReg<RA_64>(Node), TMP2.W());
+        uxtw(GetReg<RA_32>(Node), TMP2.W());
         break;
       }
       case 4: {
         aarch64::Label LoopTop;
         bind(&LoopTop);
         ldaxr(TMP2.W(), MemOperand(MemSrc));
-        stlxr(TMP4.W(), TMP3.W(), MemOperand(MemSrc));
+        stlxr(TMP4.W(), GetReg<RA_32>(Op->Header.Args[1].ID()), MemOperand(MemSrc));
         cbnz(TMP4.W(), &LoopTop);
         mov(GetReg<RA_32>(Node), TMP2.W());
         break;
@@ -553,12 +554,12 @@ DEF_OP(AtomicSwap) {
         aarch64::Label LoopTop;
         bind(&LoopTop);
         ldaxr(TMP2, MemOperand(MemSrc));
-        stlxr(TMP4, TMP3.X(), MemOperand(MemSrc));
+        stlxr(TMP4, GetReg<RA_64>(Op->Header.Args[1].ID()), MemOperand(MemSrc));
         cbnz(TMP4, &LoopTop);
         mov(GetReg<RA_64>(Node), TMP2.X());
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
 }
@@ -573,7 +574,7 @@ DEF_OP(AtomicFetchAdd) {
     case 2: ldaddalh(GetReg<RA_32>(Op->Header.Args[1].ID()), GetReg<RA_32>(Node), MemOperand(MemSrc)); break;
     case 4: ldaddal(GetReg<RA_32>(Op->Header.Args[1].ID()), GetReg<RA_32>(Node), MemOperand(MemSrc)); break;
     case 8: ldaddal(GetReg<RA_64>(Op->Header.Args[1].ID()), GetReg<RA_64>(Node), MemOperand(MemSrc)); break;
-    default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
   else {
@@ -619,7 +620,7 @@ DEF_OP(AtomicFetchAdd) {
         mov(GetReg<RA_64>(Node), TMP2);
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
 }
@@ -635,7 +636,7 @@ DEF_OP(AtomicFetchSub) {
     case 2: ldaddalh(TMP2.W(), GetReg<RA_32>(Node), MemOperand(MemSrc)); break;
     case 4: ldaddal(TMP2.W(), GetReg<RA_32>(Node), MemOperand(MemSrc)); break;
     case 8: ldaddal(TMP2.X(), GetReg<RA_64>(Node), MemOperand(MemSrc)); break;
-    default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
   else {
@@ -681,7 +682,7 @@ DEF_OP(AtomicFetchSub) {
         mov(GetReg<RA_64>(Node), TMP2);
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
 }
@@ -697,7 +698,7 @@ DEF_OP(AtomicFetchAnd) {
     case 2: ldclralh(TMP2.W(), GetReg<RA_32>(Node), MemOperand(MemSrc)); break;
     case 4: ldclral(TMP2.W(), GetReg<RA_32>(Node), MemOperand(MemSrc)); break;
     case 8: ldclral(TMP2.X(), GetReg<RA_64>(Node), MemOperand(MemSrc)); break;
-    default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
   else {
@@ -743,7 +744,7 @@ DEF_OP(AtomicFetchAnd) {
         mov(GetReg<RA_64>(Node), TMP2);
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
 }
@@ -758,7 +759,7 @@ DEF_OP(AtomicFetchOr) {
     case 2: ldsetalh(GetReg<RA_32>(Op->Header.Args[1].ID()), GetReg<RA_32>(Node), MemOperand(MemSrc)); break;
     case 4: ldsetal(GetReg<RA_32>(Op->Header.Args[1].ID()), GetReg<RA_32>(Node), MemOperand(MemSrc)); break;
     case 8: ldsetal(GetReg<RA_64>(Op->Header.Args[1].ID()), GetReg<RA_64>(Node), MemOperand(MemSrc)); break;
-    default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
   else {
@@ -804,7 +805,7 @@ DEF_OP(AtomicFetchOr) {
         mov(GetReg<RA_64>(Node), TMP2);
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
 }
@@ -819,7 +820,7 @@ DEF_OP(AtomicFetchXor) {
     case 2: ldeoralh(GetReg<RA_32>(Op->Header.Args[1].ID()), GetReg<RA_32>(Node), MemOperand(MemSrc)); break;
     case 4: ldeoral(GetReg<RA_32>(Op->Header.Args[1].ID()), GetReg<RA_32>(Node), MemOperand(MemSrc)); break;
     case 8: ldeoral(GetReg<RA_64>(Op->Header.Args[1].ID()), GetReg<RA_64>(Node), MemOperand(MemSrc)); break;
-    default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
   else {
@@ -865,11 +866,61 @@ DEF_OP(AtomicFetchXor) {
         mov(GetReg<RA_64>(Node), TMP2);
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled Atomic size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
     }
   }
 }
 
+DEF_OP(AtomicFetchNeg) {
+  auto Op = IROp->C<IR::IROp_AtomicFetchNeg>();
+  auto MemSrc = GetReg<RA_64>(Op->Header.Args[0].ID());
+
+  // TMP2-TMP3
+  switch (Op->Size) {
+    case 1: {
+      aarch64::Label LoopTop;
+      bind(&LoopTop);
+      ldaxrb(TMP2.W(), MemOperand(MemSrc));
+      neg(TMP3.W(), TMP2.W());
+      stlxrb(TMP4.W(), TMP3.W(), MemOperand(MemSrc));
+      cbnz(TMP4.W(), &LoopTop);
+      mov(GetReg<RA_32>(Node), TMP2.W());
+      break;
+    }
+    case 2: {
+      aarch64::Label LoopTop;
+      bind(&LoopTop);
+      ldaxrh(TMP2.W(), MemOperand(MemSrc));
+      neg(TMP3.W(), TMP2.W());
+      stlxrh(TMP4.W(), TMP3.W(), MemOperand(MemSrc));
+      cbnz(TMP4.W(), &LoopTop);
+      mov(GetReg<RA_32>(Node), TMP2.W());
+      break;
+    }
+    case 4: {
+      aarch64::Label LoopTop;
+      bind(&LoopTop);
+      ldaxr(TMP2.W(), MemOperand(MemSrc));
+      neg(TMP3.W(), TMP2.W());
+      stlxr(TMP4.W(), TMP3.W(), MemOperand(MemSrc));
+      cbnz(TMP4.W(), &LoopTop);
+      mov(GetReg<RA_32>(Node), TMP2.W());
+      break;
+    }
+    case 8: {
+      aarch64::Label LoopTop;
+      bind(&LoopTop);
+      ldaxr(TMP2, MemOperand(MemSrc));
+      neg(TMP3, TMP2);
+      stlxr(TMP4, TMP3, MemOperand(MemSrc));
+      cbnz(TMP4, &LoopTop);
+      mov(GetReg<RA_64>(Node), TMP2);
+      break;
+    }
+    default:  LOGMAN_MSG_A_FMT("Unhandled Atomic size: {}", Op->Size);
+  }
+}
+
 #undef DEF_OP
 void Arm64JITCore::RegisterAtomicHandlers() {
 #define REGISTER_OP(op, x) OpHandlers[FEXCore::IR::IROps::OP_##op] = &Arm64JITCore::Op_##x
@@ -886,6 +937,7 @@ void Arm64JITCore::RegisterAtomicHandlers() {
   REGISTER_OP(ATOMICFETCHAND, AtomicFetchAnd);
   REGISTER_OP(ATOMICFETCHOR,  AtomicFetchOr);
   REGISTER_OP(ATOMICFETCHXOR, AtomicFetchXor);
+  REGISTER_OP(ATOMICFETCHNEG, AtomicFetchNeg);
 #undef REGISTER_OP
 }
 }

External/FEXCore/Source/Interface/Core/JIT/Arm64/BranchOps.cpp

@@ -4,6 +4,8 @@ tags: backend|arm64
 $end_info$
 */
 
+#include "Interface/Core/LookupCache.h"
+
 #include "Interface/Core/JIT/Arm64/JITClass.h"
 #include "Interface/Core/InternalThreadState.h"
 
@@ -16,15 +18,15 @@ using namespace vixl;
 using namespace vixl::aarch64;
 #define DEF_OP(x) void Arm64JITCore::Op_##x(FEXCore::IR::IROp_Header *IROp, uint32_t Node)
 DEF_OP(GuestCallDirect) {
-  LogMan::Msg::D("Unimplemented");
+  LogMan::Msg::DFmt("Unimplemented");
 }
 
 DEF_OP(GuestCallIndirect) {
-  LogMan::Msg::D("Unimplemented");
+  LogMan::Msg::DFmt("Unimplemented");
 }
 
 DEF_OP(GuestReturn) {
-  LogMan::Msg::D("Unimplemented");
+  LogMan::Msg::DFmt("Unimplemented");
 }
 
 DEF_OP(SignalReturn) {
@@ -142,7 +144,7 @@ Condition MapBranchCC(IR::CondClassType Cond) {
   case FEXCore::IR::COND_MI:
   case FEXCore::IR::COND_PL:
   default:
-  LOGMAN_MSG_A("Unsupported compare type");
+  LOGMAN_MSG_A_FMT("Unsupported compare type");
   return Condition::nv;
   }
 }
@@ -169,10 +171,10 @@ DEF_OP(CondJump) {
   bool isConst = IsInlineConstant(Op->Cmp2, &Const);
 
   if (isConst && Const == 0 && Op->Cond.Val == FEXCore::IR::COND_EQ) {
-    LOGMAN_THROW_A(IsGPR(Op->Cmp1.ID()), "CondJump: Expected GPR");
+    LOGMAN_THROW_A_FMT(IsGPR(Op->Cmp1.ID()), "CondJump: Expected GPR");
     cbz(GRCMP(Op->Cmp1.ID()), TrueTargetLabel);
   } else if (isConst && Const == 0 && Op->Cond.Val == FEXCore::IR::COND_NEQ) {
-    LOGMAN_THROW_A(IsGPR(Op->Cmp1.ID()), "CondJump: Expected GPR");
+    LOGMAN_THROW_A_FMT(IsGPR(Op->Cmp1.ID()), "CondJump: Expected GPR");
     cbnz(GRCMP(Op->Cmp1.ID()), TrueTargetLabel);
   } else {
     if (IsGPR(Op->Cmp1.ID())) {
@@ -183,7 +185,7 @@ DEF_OP(CondJump) {
     } else if (IsFPR(Op->Cmp1.ID())) {
       fcmp(GRFCMP(Op->Cmp1.ID()), GRFCMP(Op->Cmp2.ID()));
     } else {
-      LOGMAN_MSG_A("CondJump: Expected GPR or FPR");
+      LOGMAN_MSG_A_FMT("CondJump: Expected GPR or FPR");
     }
 
     b(TrueTargetLabel, MapBranchCC(Op->Cond));

External/FEXCore/Source/Interface/Core/JIT/Arm64/ConversionOps.cpp

@@ -31,7 +31,7 @@ DEF_OP(VInsGPR) {
       ins(GetDst(Node).V2D(), Op->Index, GetReg<RA_64>(Op->Header.Args[1].ID()));
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -52,7 +52,7 @@ DEF_OP(VCastFromGPR) {
     case 8:
       fmov(GetDst(Node).D(), GetReg<RA_64>(Op->Header.Args[0].ID()).X());
       break;
-    default: LOGMAN_MSG_A("Unknown castGPR element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown castGPR element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -91,7 +91,7 @@ DEF_OP(Float_FToF) {
       fcvt(GetDst(Node).S(), GetSrc(Op->Header.Args[0].ID()).D());
       break;
     }
-    default: LOGMAN_MSG_A("Unknown FCVT sizes: 0x%x", Conv);
+    default: LOGMAN_MSG_A_FMT("Unknown FCVT sizes: 0x{:x}", Conv);
   }
 }
 
@@ -104,7 +104,7 @@ DEF_OP(Vector_SToF) {
     case 8:
       scvtf(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D());
     break;
-    default: LOGMAN_MSG_A("Unknown castGPR element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Vector_SToF element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -117,7 +117,7 @@ DEF_OP(Vector_FToZS) {
     case 8:
       fcvtzs(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D());
     break;
-    default: LOGMAN_MSG_A("Unknown castGPR element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Vector_FToZS element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -132,7 +132,7 @@ DEF_OP(Vector_FToS) {
       frinti(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D());
       fcvtzs(GetDst(Node).V2D(), GetDst(Node).V2D());
     break;
-    default: LOGMAN_MSG_A("Unknown castGPR element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Vector_FToS element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -149,7 +149,7 @@ DEF_OP(Vector_FToF) {
       fcvtn(GetDst(Node).V2S(), GetSrc(Op->Header.Args[0].ID()).V2D());
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Conversion Type : 0%04x", Conv); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Vector_FToF Type : 0x{:04x}", Conv); break;
   }
 }
 

External/FEXCore/Source/Interface/Core/JIT/Arm64/JIT.cpp

@@ -11,6 +11,7 @@ $end_info$
 */
 
 #include "Interface/Context/Context.h"
+#include "Interface/Core/LookupCache.h"
 
 #include "Interface/Core/ArchHelpers/Arm64.h"
 #include "Interface/Core/ArchHelpers/MContext.h"
@@ -45,8 +46,7 @@ void Arm64JITCore::Op_Unhandled(FEXCore::IR::IROp_Header *IROp, uint32_t Node) {
   FallbackInfo Info;
   if (!InterpreterOps::GetFallbackHandler(IROp, &Info)) {
 #if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
-    auto Name = FEXCore::IR::GetName(IROp->Op);
-    LOGMAN_MSG_A("Unhandled IR Op: %s", std::string(Name).c_str());
+    LOGMAN_MSG_A_FMT("Unhandled IR Op: {}", FEXCore::IR::GetName(IROp->Op));
 #endif
   } else {
     switch(Info.ABI) {
@@ -295,8 +295,7 @@ void Arm64JITCore::Op_Unhandled(FEXCore::IR::IROp_Header *IROp, uint32_t Node) {
       case FABI_UNKNOWN:
       default:
 #if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
-        auto Name = FEXCore::IR::GetName(IROp->Op);
-        LOGMAN_MSG_A("Unhandled IR Fallback abi: %s %d", std::string(Name).c_str(), Info.ABI);
+        LOGMAN_MSG_A_FMT("Unhandled IR Fallback ABI: {} {}", FEXCore::IR::GetName(IROp->Op), Info.ABI);
 #endif
       break;
     }
@@ -315,7 +314,7 @@ Arm64JITCore::CodeBuffer Arm64JITCore::AllocateNewCodeBuffer(size_t Size) {
                     PROT_READ | PROT_WRITE | PROT_EXEC,
                     MAP_PRIVATE | MAP_ANONYMOUS,
                     -1, 0));
-  LOGMAN_THROW_A(!!Buffer.Ptr, "Couldn't allocate code buffer");
+  LOGMAN_THROW_A_FMT(!!Buffer.Ptr, "Couldn't allocate code buffer");
   Dispatcher->RegisterCodeBuffer(Buffer.Ptr, Buffer.Size);
   return Buffer;
 }
@@ -345,54 +344,70 @@ bool Arm64JITCore::HandleSIGBUS(int Signal, void *info, void *ucontext) {
     0b1011'0000'0000; // Inner shareable all
   if ((Instr & 0x3F'FF'FC'00) == 0x08'DF'FC'00 || // LDAR*
       (Instr & 0x3F'FF'FC'00) == 0x38'BF'C0'00) { // LDAPR*
-    uint32_t LDR = 0b0011'1000'0111'1111'0110'1000'0000'0000;
-    LDR |= Size << 30;
-    LDR |= AddrReg << 5;
-    LDR |= DataReg;
-    PC[-1] = DMB;
-    PC[0] = LDR;
-    PC[1] = DMB;
-    // Back up one instruction and have another go
-    ArchHelpers::Context::SetPc(ucontext, ArchHelpers::Context::GetPc(ucontext) - 4);
+    if (ParanoidTSO()) {
+      if (FEXCore::ArchHelpers::Arm64::HandleAtomicLoad(ucontext, info, Instr)) {
+        // Skip this instruction now
+        ArchHelpers::Context::SetPc(ucontext, ArchHelpers::Context::GetPc(ucontext) + 4);
+        return true;
+      }
+      else {
+        LogMan::Msg::EFmt("Unhandled JIT SIGBUS LDAR*: PC: {} Instruction: 0x{:08x}\n", fmt::ptr(PC), PC[0]);
+        return false;
+      }
+    }
+    else {
+      uint32_t LDR = 0b0011'1000'0111'1111'0110'1000'0000'0000;
+      LDR |= Size << 30;
+      LDR |= AddrReg << 5;
+      LDR |= DataReg;
+      PC[-1] = DMB;
+      PC[0] = LDR;
+      PC[1] = DMB;
+      // Back up one instruction and have another go
+      ArchHelpers::Context::SetPc(ucontext, ArchHelpers::Context::GetPc(ucontext) - 4);
+    }
   }
   else if ( (Instr & 0x3F'FF'FC'00) == 0x08'9F'FC'00) { // STLR*
-    uint32_t STR = 0b0011'1000'0011'1111'0110'1000'0000'0000;
-    STR |= Size << 30;
-    STR |= AddrReg << 5;
-    STR |= DataReg;
-    PC[-1] = DMB;
-    PC[0] = STR;
-    PC[1] = DMB;
-    // Back up one instruction and have another go
-    ArchHelpers::Context::SetPc(ucontext, ArchHelpers::Context::GetPc(ucontext) - 4);
+    if (ParanoidTSO()) {
+      if (FEXCore::ArchHelpers::Arm64::HandleAtomicStore(ucontext, info, Instr)) {
+        // Skip this instruction now
+        ArchHelpers::Context::SetPc(ucontext, ArchHelpers::Context::GetPc(ucontext) + 4);
+        return true;
+      }
+      else {
+        LogMan::Msg::EFmt("Unhandled JIT SIGBUS STLR*: PC: {} Instruction: 0x{:08x}\n", fmt::ptr(PC), PC[0]);
+        return false;
+      }
+    }
+    else {
+      uint32_t STR = 0b0011'1000'0011'1111'0110'1000'0000'0000;
+      STR |= Size << 30;
+      STR |= AddrReg << 5;
+      STR |= DataReg;
+      PC[-1] = DMB;
+      PC[0] = STR;
+      PC[1] = DMB;
+      // Back up one instruction and have another go
+      ArchHelpers::Context::SetPc(ucontext, ArchHelpers::Context::GetPc(ucontext) - 4);
+    }
   }
   else if ((Instr & FEXCore::ArchHelpers::Arm64::LDAXP_MASK) == FEXCore::ArchHelpers::Arm64::LDAXP_INST) { // LDAXP
-    uint32_t DataReg2 = (Instr >> 10) & 0x1F;
-    // Convert to LDP
-    uint32_t LDP = 0b0010'1001'0100'0000'0000'0000'0000'0000;
-    LDP |= Size << 31;
-    LDP |= DataReg2 << 10;
-    LDP |= AddrReg << 5;
-    LDP |= DataReg;
-    PC[-1] = DMB;
-    PC[0] = LDP;
-    PC[1] = DMB;
-    // Back up one instruction and have another go
-    ArchHelpers::Context::SetPc(ucontext, ArchHelpers::Context::GetPc(ucontext) - 4);
+    //Should be compare and swap pair only. LDAXP not used elsewhere
+    uint64_t BytesToSkip = FEXCore::ArchHelpers::Arm64::HandleCASPAL_ARMv8(ucontext, info, Instr);
+    if (BytesToSkip) {
+      // Skip this instruction now
+      ArchHelpers::Context::SetPc(ucontext, ArchHelpers::Context::GetPc(ucontext) + BytesToSkip);
+      return true;
+    }
+    else {
+      LogMan::Msg::EFmt("Unhandled JIT SIGBUS LDAXP: PC: {} Instruction: 0x{:08x}\n", fmt::ptr(PC), PC[0]);
+      return false;
+    }
   }
   else if ((Instr & FEXCore::ArchHelpers::Arm64::STLXP_MASK) == FEXCore::ArchHelpers::Arm64::STLXP_INST) { // STLXP
-    uint32_t DataReg2 = (Instr >> 10) & 0x1F;
-    // Convert to STP
-    uint32_t STP = 0b0010'1001'0000'0000'0000'0000'0000'0000;
-    STP |= Size << 31;
-    STP |= DataReg2 << 10;
-    STP |= AddrReg << 5;
-    STP |= DataReg;
-    PC[-1] = DMB;
-    PC[0] = STP;
-    PC[1] = DMB;
-    // Back up one instruction and have another go
-    ArchHelpers::Context::SetPc(ucontext, ArchHelpers::Context::GetPc(ucontext) - 4);
+    //Should not trigger - middle of an LDAXP/STAXP pair.
+    LogMan::Msg::EFmt("Unhandled JIT SIGBUS STLXP: PC: {} Instruction: 0x{:08x}\n", fmt::ptr(PC), PC[0]);
+    return false;
   }
   else if ((Instr & FEXCore::ArchHelpers::Arm64::CASPAL_MASK) == FEXCore::ArchHelpers::Arm64::CASPAL_INST) { // CASPAL
     if (FEXCore::ArchHelpers::Arm64::HandleCASPAL(ucontext, info, Instr)) {
@@ -401,7 +416,7 @@ bool Arm64JITCore::HandleSIGBUS(int Signal, void *info, void *ucontext) {
       return true;
     }
     else {
-      LogMan::Msg::E("Unhandled JIT SIGBUS CASPAL: PC: %p Instruction: 0x%08x\n", PC, PC[0]);
+      LogMan::Msg::EFmt("Unhandled JIT SIGBUS CASPAL: PC: {} Instruction: 0x{:08x}\n", fmt::ptr(PC), PC[0]);
       return false;
     }
   }
@@ -412,7 +427,7 @@ bool Arm64JITCore::HandleSIGBUS(int Signal, void *info, void *ucontext) {
       return true;
     }
     else {
-      LogMan::Msg::E("Unhandled JIT SIGBUS CASAL: PC: %p Instruction: 0x%08x\n", PC, PC[0]);
+      LogMan::Msg::EFmt("Unhandled JIT SIGBUS CASAL: PC: {} Instruction: 0x{:08x}\n", fmt::ptr(PC), PC[0]);
       return false;
     }
   }
@@ -424,12 +439,24 @@ bool Arm64JITCore::HandleSIGBUS(int Signal, void *info, void *ucontext) {
     }
     else {
       uint8_t Op = (PC[0] >> 12) & 0xF;
-      LogMan::Msg::E("Unhandled JIT SIGBUS Atomic mem op 0x%02x: PC: %p Instruction: 0x%08x\n", Op, PC, PC[0]);
+      LogMan::Msg::EFmt("Unhandled JIT SIGBUS Atomic mem op 0x{:02x}: PC: {} Instruction: 0x{:08x}\n", Op, fmt::ptr(PC), PC[0]);
+      return false;
+    }
+  }
+  else if ((Instr & FEXCore::ArchHelpers::Arm64::LDAXR_MASK) == FEXCore::ArchHelpers::Arm64::LDAXR_INST) { // LDAXR*
+    uint64_t BytesToSkip = FEXCore::ArchHelpers::Arm64::HandleAtomicLoadstoreExclusive(ucontext, info);
+    if (BytesToSkip) {
+      // Skip this instruction now
+      ArchHelpers::Context::SetPc(ucontext, ArchHelpers::Context::GetPc(ucontext) + BytesToSkip);
+      return true;
+    }
+    else {
+      LogMan::Msg::EFmt("Unhandled JIT SIGBUS LDAXR: PC: {} Instruction: 0x{:08x}\n", fmt::ptr(PC), PC[0]);
       return false;
     }
   }
   else {
-    LogMan::Msg::E("Unhandled JIT SIGBUS: PC: %p Instruction: 0x%08x\n", PC, PC[0]);
+    LogMan::Msg::EFmt("Unhandled JIT SIGBUS: PC: {} Instruction: 0x{:08x}\n", fmt::ptr(PC), PC[0]);
     return false;
   }
 
@@ -445,7 +472,7 @@ Arm64JITCore::Arm64JITCore(FEXCore::Context::Context *ctx, FEXCore::Core::Intern
     DispatcherConfig config;
     config.ExitFunctionLink = reinterpret_cast<uintptr_t>(&ExitFunctionLink);
     config.ExitFunctionLinkThis = reinterpret_cast<uintptr_t>(this);
-    config.StaticRegisterAssignment = true;
+    config.StaticRegisterAssignment = ctx->Config.StaticRegisterAllocation;
 
     Dispatcher = std::make_unique<Arm64Dispatcher>(CTX, ThreadState, config);
     DispatchPtr = Dispatcher->DispatchPtr;
@@ -459,7 +486,7 @@ Arm64JITCore::Arm64JITCore(FEXCore::Context::Context *ctx, FEXCore::Core::Intern
 
   CurrentCodeBuffer = &InitialCodeBuffer;
 
-  RAPass = Thread->PassManager->GetRAPass();
+  RAPass = Thread->PassManager->GetPass<IR::RegisterAllocationPass>("RA");
 
 #if DEBUG
   Decoder.AppendVisitor(&Disasm)
@@ -524,7 +551,7 @@ Arm64JITCore::Arm64JITCore(FEXCore::Context::Context *ctx, FEXCore::Core::Intern
       return Core->Dispatcher->HandleGuestSignal(Signal, info, ucontext, GuestAction, GuestStack);
     };
 
-    for (uint32_t Signal = 0; Signal < SignalDelegator::MAX_SIGNALS; ++Signal) {
+    for (uint32_t Signal = 0; Signal <= SignalDelegator::MAX_SIGNALS; ++Signal) {
       CTX->SignalDelegation->RegisterHostSignalHandlerForGuest(Signal, GuestSignalHandler);
     }
   }
@@ -584,7 +611,7 @@ Arm64JITCore::~Arm64JITCore() {
 IR::PhysicalRegister Arm64JITCore::GetPhys(uint32_t Node) const {
   auto PhyReg = RAData->GetNodeRegister(Node);
 
-  LOGMAN_THROW_A(!PhyReg.IsInvalid(), "Couldn't Allocate register for node: ssa%d. Class: %d", Node, PhyReg.Class);
+  LOGMAN_THROW_A_FMT(!PhyReg.IsInvalid(), "Couldn't Allocate register for node: ssa{}. Class: {}", Node, PhyReg.Class);
 
   return PhyReg;
 }
@@ -598,7 +625,7 @@ aarch64::Register Arm64JITCore::GetReg<Arm64JITCore::RA_32>(uint32_t Node) const
   } else if (Reg.Class == IR::GPRClass.Val) {
     return RA64[Reg.Reg].W();
   } else {
-    LOGMAN_THROW_A(false, "Unexpected Class: %d", Reg.Class);
+    LOGMAN_THROW_A_FMT(false, "Unexpected Class: {}", Reg.Class);
   }
 
   FEX_UNREACHABLE;
@@ -613,7 +640,7 @@ aarch64::Register Arm64JITCore::GetReg<Arm64JITCore::RA_64>(uint32_t Node) const
   } else if (Reg.Class == IR::GPRClass.Val) {
     return RA64[Reg.Reg];
   } else {
-    LOGMAN_THROW_A(false, "Unexpected Class: %d", Reg.Class);
+    LOGMAN_THROW_A_FMT(false, "Unexpected Class: {}", Reg.Class);
   }
 
   FEX_UNREACHABLE;
@@ -639,7 +666,7 @@ aarch64::VRegister Arm64JITCore::GetSrc(uint32_t Node) const {
   } else if (Reg.Class == IR::FPRClass.Val) {
     return RAFPR[Reg.Reg];
   } else {
-    LOGMAN_THROW_A(false, "Unexpected Class: %d", Reg.Class);
+    LOGMAN_THROW_A_FMT(false, "Unexpected Class: {}", Reg.Class);
   }
 
   FEX_UNREACHABLE;
@@ -653,7 +680,7 @@ aarch64::VRegister Arm64JITCore::GetDst(uint32_t Node) const {
   } else if (Reg.Class == IR::FPRClass.Val) {
     return RAFPR[Reg.Reg];
   } else {
-    LOGMAN_THROW_A(false, "Unexpected Class: %d", Reg.Class);
+    LOGMAN_THROW_A_FMT(false, "Unexpected Class: {}", Reg.Class);
   }
 
   FEX_UNREACHABLE;
@@ -744,8 +771,7 @@ void *Arm64JITCore::CompileCode(uint64_t Entry, [[maybe_unused]] FEXCore::IR::IR
   // X1-X3 = Temp
   // X4-r18 = RA
 
-  auto Buffer = GetBuffer();
-  auto GuestEntry = Buffer->GetOffsetAddress<uint64_t>(GetCursorOffset());
+  auto GuestEntry = GetCursorAddress<uint64_t>();
 
  if (CTX->GetGdbServerStatus()) {
     aarch64::Label RunBlock;
@@ -791,7 +817,7 @@ void *Arm64JITCore::CompileCode(uint64_t Entry, [[maybe_unused]] FEXCore::IR::IR
     using namespace FEXCore::IR;
 #if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
     auto BlockIROp = BlockHeader->CW<FEXCore::IR::IROp_CodeBlock>();
-    LOGMAN_THROW_A(BlockIROp->Header.Op == IR::OP_CODEBLOCK, "IR type failed to be a code block");
+    LOGMAN_THROW_A_FMT(BlockIROp->Header.Op == IR::OP_CODEBLOCK, "IR type failed to be a code block");
 #endif
 
     {
@@ -812,7 +838,7 @@ void *Arm64JITCore::CompileCode(uint64_t Entry, [[maybe_unused]] FEXCore::IR::IR
     }
 
     if (DebugData) {
-      DebugData->Subblocks.push_back({Buffer->GetOffsetAddress<uintptr_t>(GetCursorOffset()), 0, IR->GetID(BlockNode)});
+      DebugData->Subblocks.push_back({GetCursorAddress<uintptr_t>(), 0, IR->GetID(BlockNode)});
     }
 
     for (auto [CodeNode, IROp] : IR->GetCode(BlockNode)) {
@@ -824,7 +850,7 @@ void *Arm64JITCore::CompileCode(uint64_t Entry, [[maybe_unused]] FEXCore::IR::IR
     }
 
     if (DebugData) {
-      DebugData->Subblocks.back().HostCodeSize = Buffer->GetOffsetAddress<uintptr_t>(GetCursorOffset()) - DebugData->Subblocks.back().HostCodeStart;
+      DebugData->Subblocks.back().HostCodeSize = GetCursorAddress<uintptr_t>() - DebugData->Subblocks.back().HostCodeStart;
     }
   }
 
@@ -837,7 +863,7 @@ void *Arm64JITCore::CompileCode(uint64_t Entry, [[maybe_unused]] FEXCore::IR::IR
 
   FinalizeCode();
 
-  auto CodeEnd = Buffer->GetOffsetAddress<uint64_t>(GetCursorOffset());
+  auto CodeEnd = GetCursorAddress<uint64_t>();
   CPU.EnsureIAndDCacheCoherency(reinterpret_cast<void*>(GuestEntry), CodeEnd - reinterpret_cast<uint64_t>(GuestEntry));
 
   if (DebugData) {
@@ -856,7 +882,7 @@ uint64_t Arm64JITCore::ExitFunctionLink(Arm64JITCore *core, FEXCore::Core::CpuSt
   auto HostCode = Thread->LookupCache->FindBlock(GuestRip);
 
   if (!HostCode) {
-    //printf("ExitFunctionLink: Aborting, %lX not in cache\n", GuestRip);
+    //fmt::print("ExitFunctionLink: Aborting, {:X} not in cache\n", GuestRip);
     Frame->State.rip = GuestRip;
     return core->ThreadSharedData.Dispatcher->AbsoluteLoopTopAddress;
   }

External/FEXCore/Source/Interface/Core/JIT/Arm64/JITClass.h

@@ -6,7 +6,6 @@ $end_info$
 
 #pragma once
 
-#include "Interface/Core/LookupCache.h"
 #include "Interface/Core/ArchHelpers/Arm64Emitter.h"
 #include "Interface/Core/Dispatcher/Dispatcher.h"
 
@@ -257,6 +256,7 @@ private:
   DEF_OP(AtomicFetchAnd);
   DEF_OP(AtomicFetchOr);
   DEF_OP(AtomicFetchXor);
+  DEF_OP(AtomicFetchNeg);
 
   ///< Branch ops
   DEF_OP(GuestCallDirect);

External/FEXCore/Source/Interface/Core/JIT/Arm64/MemoryOps.cpp

@@ -30,7 +30,7 @@ DEF_OP(LoadContext) {
     case 8:
       ldr(GetReg<RA_64>(Node), MemOperand(STATE, Op->Offset));
     break;
-    default:  LOGMAN_MSG_A("Unhandled LoadContext size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled LoadContext size: {}", OpSize);
     }
   }
   else {
@@ -51,7 +51,7 @@ DEF_OP(LoadContext) {
     case 16:
       ldr(Dst, MemOperand(STATE, Op->Offset));
     break;
-    default:  LOGMAN_MSG_A("Unhandled LoadContext size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled LoadContext size: {}", OpSize);
     }
   }
 }
@@ -73,7 +73,7 @@ DEF_OP(StoreContext) {
     case 8:
       str(GetReg<RA_64>(Op->Header.Args[0].ID()), MemOperand(STATE, Op->Offset));
     break;
-    default:  LOGMAN_MSG_A("Unhandled StoreContext size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled StoreContext size: {}", OpSize);
     }
   }
   else {
@@ -94,7 +94,7 @@ DEF_OP(StoreContext) {
     case 16:
       str(Src, MemOperand(STATE, Op->Offset));
     break;
-    default:  LOGMAN_MSG_A("Unhandled LoadContext size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled StoreContext size: {}", OpSize);
     }
   }
 }
@@ -107,29 +107,29 @@ DEF_OP(LoadRegister) {
     auto regId = (Op->Offset - offsetof(FEXCore::Core::CpuStateFrame, State.gregs[0])) / 8;
     auto regOffs = Op->Offset & 7;
 
-    LOGMAN_THROW_A(regId < SRA64.size(), "out of range regId");
+    LOGMAN_THROW_A_FMT(regId < SRA64.size(), "out of range regId");
 
     auto reg = SRA64[regId];
 
     switch(Op->Header.Size) {
       case 1:
-        LOGMAN_THROW_A(regOffs == 0 || regOffs == 1, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT(regOffs == 0 || regOffs == 1, "unexpected regOffs");
         ubfx(GetReg<RA_64>(Node), reg, regOffs * 8, 8);
         break;
 
       case 2:
-        LOGMAN_THROW_A(regOffs == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT(regOffs == 0, "unexpected regOffs");
         ubfx(GetReg<RA_64>(Node), reg, 0, 16);
         break;
 
       case 4:
-        LOGMAN_THROW_A(regOffs == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT(regOffs == 0, "unexpected regOffs");
         if (GetReg<RA_64>(Node).GetCode() != reg.GetCode())
           mov(GetReg<RA_32>(Node), reg.W());
         break;
 
       case 8:
-        LOGMAN_THROW_A(regOffs == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT(regOffs == 0, "unexpected regOffs");
         if (GetReg<RA_64>(Node).GetCode() != reg.GetCode())
           mov(GetReg<RA_64>(Node), reg);
         break;
@@ -138,24 +138,24 @@ DEF_OP(LoadRegister) {
     auto regId = (Op->Offset - offsetof(FEXCore::Core::CpuStateFrame, State.xmm[0][0])) / 16;
     auto regOffs = Op->Offset & 15;
 
-    LOGMAN_THROW_A(regId < SRAFPR.size(), "out of range regId");
+    LOGMAN_THROW_A_FMT(regId < SRAFPR.size(), "out of range regId");
 
     auto guest = SRAFPR[regId];
     auto host = GetSrc(Node);
 
     switch(Op->Header.Size) {
       case 1:
-        LOGMAN_THROW_A(regOffs == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT(regOffs == 0, "unexpected regOffs");
         mov(host.B(), guest.B());
         break;
 
       case 2:
-        LOGMAN_THROW_A(regOffs == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT(regOffs == 0, "unexpected regOffs");
         fmov(host.H(), guest.H());
         break;
 
       case 4:
-        LOGMAN_THROW_A((regOffs & 3) == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT((regOffs & 3) == 0, "unexpected regOffs");
         if (regOffs == 0) {
           if (host.GetCode() != guest.GetCode())
             fmov(host.S(), guest.S());
@@ -165,7 +165,7 @@ DEF_OP(LoadRegister) {
         break;
 
       case 8:
-        LOGMAN_THROW_A((regOffs & 7) == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT((regOffs & 7) == 0, "unexpected regOffs");
         if (regOffs == 0) {
           if (host.GetCode() != guest.GetCode())
             mov(host.D(), guest.D());
@@ -175,13 +175,13 @@ DEF_OP(LoadRegister) {
         break;
 
       case 16:
-        LOGMAN_THROW_A(regOffs == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT(regOffs == 0, "unexpected regOffs");
         if (host.GetCode() != guest.GetCode())
           mov(host.Q(), guest.Q());
         break;
     }
   } else {
-    LOGMAN_THROW_A(false, "Unhandled Op->Class %d", Op->Class);
+    LOGMAN_THROW_A_FMT(false, "Unhandled Op->Class {}", Op->Class);
   }
 }
 
@@ -192,28 +192,28 @@ DEF_OP(StoreRegister) {
     auto regId = Op->Offset / 8 - 1;
     auto regOffs = Op->Offset & 7;
 
-    LOGMAN_THROW_A(regId < SRA64.size(), "out of range regId");
+    LOGMAN_THROW_A_FMT(regId < SRA64.size(), "out of range regId");
 
     auto reg = SRA64[regId];
 
     switch(Op->Header.Size) {
       case 1:
-        LOGMAN_THROW_A(regOffs == 0 || regOffs == 1, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT(regOffs == 0 || regOffs == 1, "unexpected regOffs");
         bfi(reg, GetReg<RA_64>(Op->Value.ID()), regOffs * 8, 8);
         break;
 
       case 2:
-        LOGMAN_THROW_A(regOffs == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT(regOffs == 0, "unexpected regOffs");
         bfi(reg, GetReg<RA_64>(Op->Value.ID()), 0, 16);
         break;
 
       case 4:
-        LOGMAN_THROW_A(regOffs == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT(regOffs == 0, "unexpected regOffs");
         bfi(reg, GetReg<RA_64>(Op->Value.ID()), 0, 32);
         break;
 
       case 8:
-        LOGMAN_THROW_A(regOffs == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT(regOffs == 0, "unexpected regOffs");
         if (GetReg<RA_64>(Op->Value.ID()).GetCode() != reg.GetCode())
           mov(reg, GetReg<RA_64>(Op->Value.ID()));
         break;
@@ -222,7 +222,7 @@ DEF_OP(StoreRegister) {
     auto regId = (Op->Offset - offsetof(FEXCore::Core::CpuStateFrame, State.xmm[0][0])) / 16;
     auto regOffs = Op->Offset & 15;
 
-    LOGMAN_THROW_A(regId < SRAFPR.size(), "regId out of range");
+    LOGMAN_THROW_A_FMT(regId < SRAFPR.size(), "regId out of range");
 
     auto guest = SRAFPR[regId];
     auto host = GetSrc(Op->Value.ID());
@@ -233,28 +233,28 @@ DEF_OP(StoreRegister) {
         break;
 
       case 2:
-        LOGMAN_THROW_A((regOffs & 1) == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT((regOffs & 1) == 0, "unexpected regOffs");
         ins(guest.V8H(), regOffs/2, host.V8H(), 0);
         break;
 
       case 4:
-        LOGMAN_THROW_A((regOffs & 3) == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT((regOffs & 3) == 0, "unexpected regOffs");
         ins(guest.V4S(), regOffs/4, host.V4S(), 0);
         break;
 
       case 8:
-        LOGMAN_THROW_A((regOffs & 7) == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT((regOffs & 7) == 0, "unexpected regOffs");
         ins(guest.V2D(), regOffs / 8, host.V2D(), 0);
         break;
 
       case 16:
-        LOGMAN_THROW_A(regOffs == 0, "unexpected regOffs");
+        LOGMAN_THROW_A_FMT(regOffs == 0, "unexpected regOffs");
         if (guest.GetCode() != host.GetCode())
           mov(guest.Q(), host.Q());
         break;
     }
   } else {
-    LOGMAN_THROW_A(false, "Unhandled Op->Class %d", Op->Class);
+    LOGMAN_THROW_A_FMT(false, "Unhandled Op->Class {}", Op->Class);
   }
 }
 
@@ -288,15 +288,17 @@ DEF_OP(LoadContextIndexed) {
         ldr(GetReg<RA_64>(Node), MemOperand(TMP1, Op->BaseOffset));
         break;
       default:
-        LOGMAN_MSG_A("Unhandled LoadContextIndexed size: %d", Op->Size);
+        LOGMAN_MSG_A_FMT("Unhandled LoadContextIndexed size: {}", Op->Size);
+        break;
       }
       break;
     }
     case 16:
-      LOGMAN_MSG_A("Invalid Class load of size 16");
+      LOGMAN_MSG_A_FMT("Invalid Class load of size 16");
       break;
     default:
-      LOGMAN_MSG_A("Unhandled LoadContextIndexed stride: %d", Op->Stride);
+      LOGMAN_MSG_A_FMT("Unhandled LoadContextIndexed stride: {}", Op->Stride);
+      break;
     }
   }
   else {
@@ -333,12 +335,14 @@ DEF_OP(LoadContextIndexed) {
         }
         break;
       default:
-        LOGMAN_MSG_A("Unhandled LoadContextIndexed size: %d", Op->Size);
+        LOGMAN_MSG_A_FMT("Unhandled LoadContextIndexed size: {}", Op->Size);
+        break;
       }
       break;
     }
     default:
-      LOGMAN_MSG_A("Unhandled LoadContextIndexed stride: %d", Op->Stride);
+      LOGMAN_MSG_A_FMT("Unhandled LoadContextIndexed stride: {}", Op->Stride);
+      break;
     }
   }
 }
@@ -374,15 +378,17 @@ DEF_OP(StoreContextIndexed) {
         str(value, MemOperand(TMP1, Op->BaseOffset));
         break;
       default:
-        LOGMAN_MSG_A("Unhandled LoadContextIndexed size: %d", Op->Size);
+        LOGMAN_MSG_A_FMT("Unhandled StoreContextIndexed size: {}", Op->Size);
+        break;
       }
       break;
     }
     case 16:
-      LOGMAN_MSG_A("Invalid Class load of size 16");
+      LOGMAN_MSG_A_FMT("Invalid Class store of size 16");
       break;
     default:
-      LOGMAN_MSG_A("Unhandled LoadContextIndexed stride: %d", Op->Stride);
+      LOGMAN_MSG_A_FMT("Unhandled StoreContextIndexed stride: {}", Op->Stride);
+      break;
     }
   }
   else {
@@ -421,12 +427,14 @@ DEF_OP(StoreContextIndexed) {
         }
         break;
       default:
-        LOGMAN_MSG_A("Unhandled LoadContextIndexed size: %d", Op->Size);
+        LOGMAN_MSG_A_FMT("Unhandled StoreContextIndexed size: {}", Op->Size);
+        break;
       }
       break;
     }
     default:
-      LOGMAN_MSG_A("Unhandled LoadContextIndexed stride: %d", Op->Stride);
+      LOGMAN_MSG_A_FMT("Unhandled StoreContextIndexed stride: {}", Op->Stride);
+      break;
     }
   }
 }
@@ -434,7 +442,7 @@ DEF_OP(StoreContextIndexed) {
 DEF_OP(SpillRegister) {
   auto Op = IROp->C<IR::IROp_SpillRegister>();
   uint8_t OpSize = IROp->Size;
-  uint32_t SlotOffset = Op->Slot * 16 + 16;
+  uint32_t SlotOffset = Op->Slot * 16;
 
   if (Op->Class == FEXCore::IR::GPRClass) {
     switch (OpSize) {
@@ -454,7 +462,7 @@ DEF_OP(SpillRegister) {
       str(GetReg<RA_64>(Op->Header.Args[0].ID()), MemOperand(sp, SlotOffset));
       break;
     }
-    default:  LOGMAN_MSG_A("Unhandled SpillRegister size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled SpillRegister size: {}", OpSize);
     }
   } else if (Op->Class == FEXCore::IR::FPRClass) {
     switch (OpSize) {
@@ -470,17 +478,17 @@ DEF_OP(SpillRegister) {
       str(GetSrc(Op->Header.Args[0].ID()), MemOperand(sp, SlotOffset));
       break;
     }
-    default:  LOGMAN_MSG_A("Unhandled SpillRegister size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled SpillRegister size: {}", OpSize);
     }
   } else {
-    LOGMAN_MSG_A("Unhandled SpillRegister class: %d", Op->Class.Val);
+    LOGMAN_MSG_A_FMT("Unhandled SpillRegister class: {}", Op->Class.Val);
   }
 }
 
 DEF_OP(FillRegister) {
   auto Op = IROp->C<IR::IROp_FillRegister>();
   uint8_t OpSize = IROp->Size;
-  uint32_t SlotOffset = Op->Slot * 16 + 16;
+  uint32_t SlotOffset = Op->Slot * 16;
 
   if (Op->Class == FEXCore::IR::GPRClass) {
     switch (OpSize) {
@@ -500,7 +508,7 @@ DEF_OP(FillRegister) {
       ldr(GetReg<RA_64>(Node), MemOperand(sp, SlotOffset));
       break;
     }
-    default:  LOGMAN_MSG_A("Unhandled SpillRegister size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled FillRegister size: {}", OpSize);
     }
   } else if (Op->Class == FEXCore::IR::FPRClass) {
     switch (OpSize) {
@@ -516,10 +524,10 @@ DEF_OP(FillRegister) {
       ldr(GetDst(Node), MemOperand(sp, SlotOffset));
       break;
     }
-    default:  LOGMAN_MSG_A("Unhandled SpillRegister size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled FillRegister size: {}", OpSize);
     }
   } else {
-    LOGMAN_MSG_A("Unhandled FillRegister class: %d", Op->Class.Val);
+    LOGMAN_MSG_A_FMT("Unhandled FillRegister class: {}", Op->Class.Val);
   }
 }
 
@@ -539,7 +547,7 @@ MemOperand Arm64JITCore::GenerateMemOperand(uint8_t AccessSize, aarch64::Registe
     return MemOperand(Base);
   } else {
     if (OffsetScale != 1 && OffsetScale != AccessSize) {
-        LOGMAN_MSG_A("Unhandled GenerateMemOperand OffsetScale: %d", OffsetScale);
+        LOGMAN_MSG_A_FMT("Unhandled GenerateMemOperand OffsetScale: {}", OffsetScale);
     }
     uint64_t Const;
     if (IsInlineConstant(Offset, &Const)) {
@@ -551,7 +559,7 @@ MemOperand Arm64JITCore::GenerateMemOperand(uint8_t AccessSize, aarch64::Registe
         case IR::MEM_OFFSET_UXTW.Val: return MemOperand(Base, RegOffset.W(), Extend::UXTW, (int)std::log2(OffsetScale) );
         case IR::MEM_OFFSET_SXTW.Val: return MemOperand(Base, RegOffset.W(), Extend::SXTW, (int)std::log2(OffsetScale) );
 
-        default: LOGMAN_MSG_A("Unhandled GenerateMemOperand OffsetType: %d", OffsetType.Val); break;
+        default: LOGMAN_MSG_A_FMT("Unhandled GenerateMemOperand OffsetType: {}", OffsetType.Val); break;
       }
     }
   }
@@ -580,7 +588,7 @@ DEF_OP(LoadMem) {
       case 8:
         ldr(Dst, MemSrc);
         break;
-      default:  LOGMAN_MSG_A("Unhandled LoadMem size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled LoadMem size: {}", Op->Size);
     }
   }
   else {
@@ -601,7 +609,7 @@ DEF_OP(LoadMem) {
       case 16:
         ldr(Dst, MemSrc);
         break;
-      default:  LOGMAN_MSG_A("Unhandled LoadMem size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled LoadMem size: {}", Op->Size);
     }
   }
 }
@@ -612,7 +620,7 @@ DEF_OP(LoadMemTSO) {
   auto MemSrc = MemOperand(GetReg<RA_64>(Op->Header.Args[0].ID()));
 
   if (!Op->Offset.IsInvalid()) {
-    LOGMAN_MSG_A("LoadMemTSO: No offset allowed");
+    LOGMAN_MSG_A_FMT("LoadMemTSO: No offset allowed");
   }
 
   if (SupportsRCPC && Op->Class == FEXCore::IR::GPRClass) {
@@ -635,7 +643,7 @@ DEF_OP(LoadMemTSO) {
         case 8:
           ldapr(Dst, MemSrc);
           break;
-        default:  LOGMAN_MSG_A("Unhandled LoadMem size: %d", Op->Size);
+        default:  LOGMAN_MSG_A_FMT("Unhandled LoadMemTSO size: {}", Op->Size);
       }
       nop();
     }
@@ -660,7 +668,7 @@ DEF_OP(LoadMemTSO) {
         case 8:
           ldar(Dst, MemSrc);
           break;
-        default:  LOGMAN_MSG_A("Unhandled LoadMem size: %d", Op->Size);
+        default:  LOGMAN_MSG_A_FMT("Unhandled LoadMemTSO size: {}", Op->Size);
       }
       nop();
     }
@@ -681,7 +689,7 @@ DEF_OP(LoadMemTSO) {
       case 16:
         ldr(Dst, MemSrc);
         break;
-      default:  LOGMAN_MSG_A("Unhandled LoadMem size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled LoadMemTSO size: {}", Op->Size);
     }
     dmb(InnerShareable, BarrierAll);
   }
@@ -708,7 +716,7 @@ DEF_OP(StoreMem) {
       case 8:
         str(GetReg<RA_64>(Op->Header.Args[1].ID()), MemSrc);
         break;
-      default:  LOGMAN_MSG_A("Unhandled StoreMem size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled StoreMem size: {}", Op->Size);
     }
   }
   else {
@@ -729,7 +737,7 @@ DEF_OP(StoreMem) {
       case 16:
         str(Src, MemSrc);
         break;
-      default:  LOGMAN_MSG_A("Unhandled StoreMem size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled StoreMem size: {}", Op->Size);
     }
   }
 }
@@ -739,7 +747,7 @@ DEF_OP(StoreMemTSO) {
   auto MemSrc = MemOperand(GetReg<RA_64>(Op->Header.Args[0].ID()));
 
   if (!Op->Offset.IsInvalid()) {
-    LOGMAN_MSG_A("StoreMemTSO: No offset allowed");
+    LOGMAN_MSG_A_FMT("StoreMemTSO: No offset allowed");
   }
 
   if (Op->Class == FEXCore::IR::GPRClass) {
@@ -759,7 +767,7 @@ DEF_OP(StoreMemTSO) {
         case 8:
           stlr(GetReg<RA_64>(Op->Header.Args[1].ID()), MemSrc);
           break;
-        default:  LOGMAN_MSG_A("Unhandled StoreMem size: %d", Op->Size);
+        default:  LOGMAN_MSG_A_FMT("Unhandled StoreMemTSO size: {}", Op->Size);
       }
       nop();
     }
@@ -783,7 +791,7 @@ DEF_OP(StoreMemTSO) {
       case 16:
         str(Src, MemSrc);
         break;
-      default:  LOGMAN_MSG_A("Unhandled StoreMem size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled StoreMemTSO size: {}", Op->Size);
     }
     dmb(InnerShareable, BarrierAll);
   }
@@ -795,7 +803,7 @@ DEF_OP(ParanoidLoadMemTSO) {
   auto MemSrc = MemOperand(GetReg<RA_64>(Op->Header.Args[0].ID()));
 
   if (!Op->Offset.IsInvalid()) {
-    LOGMAN_MSG_A("LoadMemTSO: No offset allowed");
+    LOGMAN_MSG_A_FMT("ParanoidLoadMemTSO: No offset allowed");
   }
 
   if (Op->Class == FEXCore::IR::GPRClass) {
@@ -806,7 +814,6 @@ DEF_OP(ParanoidLoadMemTSO) {
     }
     else {
       auto Dst = GetReg<RA_64>(Node);
-      nop();
       switch (Op->Size) {
         case 2:
           ldarh(Dst, MemSrc);
@@ -817,31 +824,24 @@ DEF_OP(ParanoidLoadMemTSO) {
         case 8:
           ldar(Dst, MemSrc);
           break;
-        default:  LOGMAN_MSG_A("Unhandled LoadMem size: %d", Op->Size);
+        default:  LOGMAN_MSG_A_FMT("Unhandled ParanoidLoadMemTSO size: {}", Op->Size);
       }
-      nop();
     }
   }
   else {
     auto Dst = GetDst(Node);
     switch (Op->Size) {
       case 2:
-        nop();
-        ldarh(TMP1, MemSrc);
-        nop();
-        fmov(Dst, TMP1);
+        ldarh(TMP1.W(), MemSrc);
+        fmov(Dst.H(), TMP1.W());
         break;
       case 4:
-        nop();
         ldar(TMP1.W(), MemSrc);
-        nop();
-        fmov(Dst, TMP1);
+        fmov(Dst.S(), TMP1.W());
         break;
       case 8:
-        nop();
         ldar(TMP1, MemSrc);
-        nop();
-        fmov(Dst, TMP1);
+        fmov(Dst.D(), TMP1);
         break;
       case 16:
         nop();
@@ -850,7 +850,7 @@ DEF_OP(ParanoidLoadMemTSO) {
         mov(Dst.V2D(), 0, TMP1);
         mov(Dst.V2D(), 1, TMP2);
         break;
-      default:  LOGMAN_MSG_A("Unhandled LoadMem size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled ParanoidLoadMemTSO size: {}", Op->Size);
     }
   }
 }
@@ -860,7 +860,7 @@ DEF_OP(ParanoidStoreMemTSO) {
   auto MemSrc = MemOperand(GetReg<RA_64>(Op->Header.Args[0].ID()));
 
   if (!Op->Offset.IsInvalid()) {
-    LOGMAN_MSG_A("StoreMemTSO: No offset allowed");
+    LOGMAN_MSG_A_FMT("ParanoidStoreMemTSO: No offset allowed");
   }
 
   if (Op->Class == FEXCore::IR::GPRClass) {
@@ -869,7 +869,6 @@ DEF_OP(ParanoidStoreMemTSO) {
       stlrb(GetReg<RA_64>(Op->Header.Args[1].ID()), MemSrc);
     }
     else {
-      nop();
       switch (Op->Size) {
         case 2:
           stlrh(GetReg<RA_64>(Op->Header.Args[1].ID()), MemSrc);
@@ -880,37 +879,30 @@ DEF_OP(ParanoidStoreMemTSO) {
         case 8:
           stlr(GetReg<RA_64>(Op->Header.Args[1].ID()), MemSrc);
           break;
-        default:  LOGMAN_MSG_A("Unhandled StoreMem size: %d", Op->Size);
+        default:  LOGMAN_MSG_A_FMT("Unhandled ParanoidStoreMemTSO size: {}", Op->Size);
       }
-      nop();
     }
   }
   else {
     auto Src = GetSrc(Op->Header.Args[1].ID());
     if (Op->Size == 1) {
       // 8bit load is always aligned to natural alignment
-      mov(TMP1, Src.V4S(), 0);
+      mov(TMP1.W(), Src.V16B(), 0);
       stlrb(TMP1, MemSrc);
     }
     else {
       switch (Op->Size) {
         case 2:
-          mov(TMP1, Src.V4S(), 0);
-          nop();
+          mov(TMP1.W(), Src.V8H(), 0);
           stlrh(TMP1, MemSrc);
-          nop();
           break;
         case 4:
-          mov(TMP1, Src.V4S(), 0);
-          nop();
+          mov(TMP1.W(), Src.V4S(), 0);
           stlr(TMP1.W(), MemSrc);
-          nop();
           break;
         case 8:
           mov(TMP1, Src.V2D(), 0);
-          nop();
           stlr(TMP1, MemSrc);
-          nop();
           break;
         case 16: {
           // Move vector to GPRs
@@ -927,18 +919,18 @@ DEF_OP(ParanoidStoreMemTSO) {
           cbnz(TMP3, &B); // < Overwritten with DMB
           break;
         }
-        default:  LOGMAN_MSG_A("Unhandled StoreMem size: %d", Op->Size);
+        default:  LOGMAN_MSG_A_FMT("Unhandled ParanoidStoreMemTSO size: {}", Op->Size);
       }
     }
   }
 }
 
 DEF_OP(VLoadMemElement) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(VStoreMemElement) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(CacheLineClear) {

External/FEXCore/Source/Interface/Core/JIT/Arm64/MiscOps.cpp

@@ -8,11 +8,11 @@ $end_info$
 
 namespace FEXCore::CPU {
 static void PrintValue(uint64_t Value) {
-  LogMan::Msg::D("Value: 0x%lx", Value);
+  LogMan::Msg::DFmt("Value: 0x{:x}", Value);
 }
 
 static void PrintVectorValue(uint64_t Value, uint64_t ValueUpper) {
-  LogMan::Msg::D("Value: 0x%016lx'%016lx", ValueUpper, Value);
+  LogMan::Msg::DFmt("Value: 0x{:016x}'{:016x}", ValueUpper, Value);
 }
 
 using namespace vixl;
@@ -31,7 +31,7 @@ DEF_OP(Fence) {
     case IR::Fence_Store.Val:
       dmb(FullSystem, BarrierWrites);
       break;
-    default: LOGMAN_MSG_A("Unknown Fence: %d", Op->Fence); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Fence: {}", Op->Fence); break;
   }
 }
 
@@ -69,7 +69,7 @@ DEF_OP(Break) {
       br(TMP1);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Break reason: %d", Op->Reason);
+    default: LOGMAN_MSG_A_FMT("Unknown Break reason: {}", Op->Reason);
   }
 }
 

External/FEXCore/Source/Interface/Core/JIT/Arm64/MoveOps.cpp

@@ -26,7 +26,7 @@ DEF_OP(ExtractElementPair) {
       mov (GetReg<RA_64>(Node), Regs[Op->Element]);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Size"); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Size"); break;
   }
 }
 
@@ -52,7 +52,7 @@ DEF_OP(CreateElementPair) {
       RegTmp = x0;
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Size"); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Size"); break;
   }
 
   if (Dst.first.GetCode() != RegSecond.GetCode()) {

External/FEXCore/Source/Interface/Core/JIT/Arm64/VectorOps.cpp

@@ -22,7 +22,7 @@ DEF_OP(VectorZero) {
        eor(GetDst(Node).V16B(), GetDst(Node).V16B(), GetDst(Node).V16B());
        break;
      }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", OpSize); break;
   }
 }
 
@@ -43,47 +43,47 @@ DEF_OP(VectorImm) {
 }
 
 DEF_OP(CreateVector2) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(CreateVector4) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(SplatVector2) {
-	auto Op = IROp->C<IR::IROp_SplatVector2>();
+  auto Op = IROp->C<IR::IROp_SplatVector2>();
   uint8_t OpSize = IROp->Size;
-	LOGMAN_THROW_A(OpSize <= 16, "Can't handle a vector of size: %d", OpSize);
+  LOGMAN_THROW_A_FMT(OpSize <= 16, "Can't handle a vector of size: {}", OpSize);
 
-	uint8_t ElementSize = OpSize / 2;
+  uint8_t ElementSize = OpSize / 2;
 
-	switch (ElementSize) {
-		case 4:
-			dup(GetDst(Node).V4S(), GetSrc(Op->Header.Args[0].ID()).V4S(), 0);
-		break;
-		case 8:
-			dup(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), 0);
-		break;
-		default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.Size); break;
-	}
+  switch (ElementSize) {
+    case 4:
+      dup(GetDst(Node).V4S(), GetSrc(Op->Header.Args[0].ID()).V4S(), 0);
+    break;
+    case 8:
+      dup(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), 0);
+    break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.Size); break;
+  }
 }
 
 DEF_OP(SplatVector4) {
-	auto Op = IROp->C<IR::IROp_SplatVector4>();
+  auto Op = IROp->C<IR::IROp_SplatVector4>();
   uint8_t OpSize = IROp->Size;
-	LOGMAN_THROW_A(OpSize <= 16, "Can't handle a vector of size: %d", OpSize);
+  LOGMAN_THROW_A_FMT(OpSize <= 16, "Can't handle a vector of size: {}", OpSize);
 
-	uint8_t ElementSize = OpSize / 4;
+  uint8_t ElementSize = OpSize / 4;
 
-	switch (ElementSize) {
-		case 4:
-			dup(GetDst(Node).V4S(), GetSrc(Op->Header.Args[0].ID()).V4S(), 0);
-		break;
-		case 8:
-			dup(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), 0);
-		break;
-		default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.Size); break;
-	}
+  switch (ElementSize) {
+    case 4:
+      dup(GetDst(Node).V4S(), GetSrc(Op->Header.Args[0].ID()).V4S(), 0);
+    break;
+    case 8:
+      dup(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), 0);
+    break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.Size); break;
+  }
 }
 
 DEF_OP(VMov) {
@@ -118,7 +118,7 @@ DEF_OP(VMov) {
 			  mov(GetDst(Node).V16B(), GetSrc(Op->Header.Args[0].ID()).V16B());
 			break;
 		}
-		default: LOGMAN_MSG_A("Unknown Element Size: %d", OpSize); break;
+		default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", OpSize); break;
 	}
 }
 
@@ -161,7 +161,7 @@ DEF_OP(VAdd) {
       add(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -184,7 +184,7 @@ DEF_OP(VSub) {
       sub(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -207,7 +207,7 @@ DEF_OP(VUQAdd) {
       uqadd(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -230,7 +230,7 @@ DEF_OP(VUQSub) {
       uqsub(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -253,7 +253,7 @@ DEF_OP(VSQAdd) {
       sqadd(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -276,7 +276,7 @@ DEF_OP(VSQSub) {
       sqsub(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -298,7 +298,7 @@ DEF_OP(VAddP) {
         addp(GetDst(Node).V2S(), GetSrc(Op->Header.Args[0].ID()).V2S(), GetSrc(Op->Header.Args[1].ID()).V2S());
         break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -319,7 +319,7 @@ DEF_OP(VAddP) {
         addp(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
         break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -338,7 +338,7 @@ DEF_OP(VAddV) {
     case 8:
       addp(GetDst(Node).VCast(OpSize * 8, 1), GetSrc(Op->Header.Args[0].ID()).VCast(OpSize * 8, Elements));
       break;
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -353,7 +353,7 @@ DEF_OP(VUMinV) {
     case 4:
       uminv(GetDst(Node).VCast(Op->Header.ElementSize * 8, 1), GetSrc(Op->Header.Args[0].ID()).VCast(OpSize * 8, Elements));
       break;
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -368,7 +368,7 @@ DEF_OP(VURAvg) {
       urhadd(GetDst(Node).V8H(), GetSrc(Op->Header.Args[0].ID()).V8H(), GetSrc(Op->Header.Args[1].ID()).V8H());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -383,7 +383,7 @@ DEF_OP(VAbs) {
         abs(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -395,7 +395,7 @@ DEF_OP(VAbs) {
       case 8:
         abs(GetDst(Node).VCast(OpSize * 8, Elements), GetSrc(Op->Header.Args[0].ID()).VCast(OpSize * 8, Elements));
         break;
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -410,7 +410,7 @@ DEF_OP(VPopcount) {
         cnt(GetDst(Node).V8B(), GetSrc(Op->Header.Args[0].ID()).V8B());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -419,7 +419,7 @@ DEF_OP(VPopcount) {
       case 1:
         cnt(GetDst(Node).V16B(), GetSrc(Op->Header.Args[0].ID()).V16B());
         break;
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -438,7 +438,7 @@ DEF_OP(VFAdd) {
         fadd(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), GetSrc(Op->Header.Args[1].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -452,7 +452,7 @@ DEF_OP(VFAdd) {
         fadd(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -468,7 +468,7 @@ DEF_OP(VFAddP) {
       faddp(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -486,7 +486,7 @@ DEF_OP(VFSub) {
         fsub(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), GetSrc(Op->Header.Args[1].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -500,7 +500,7 @@ DEF_OP(VFSub) {
         fsub(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -519,7 +519,7 @@ DEF_OP(VFMul) {
         fmul(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), GetSrc(Op->Header.Args[1].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -533,7 +533,7 @@ DEF_OP(VFMul) {
         fmul(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -552,7 +552,7 @@ DEF_OP(VFDiv) {
         fdiv(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), GetSrc(Op->Header.Args[1].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -566,7 +566,7 @@ DEF_OP(VFDiv) {
         fdiv(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -587,7 +587,7 @@ DEF_OP(VFMin) {
         fcsel(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), GetSrc(Op->Header.Args[1].ID()).D(), Condition::mi);
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -607,7 +607,7 @@ DEF_OP(VFMin) {
         mov(GetDst(Node).V2D(), VTMP2.V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -628,7 +628,7 @@ DEF_OP(VFMax) {
         fcsel(GetDst(Node).D(), GetSrc(Op->Header.Args[1].ID()).D(), GetSrc(Op->Header.Args[0].ID()).D(), Condition::mi);
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -648,7 +648,7 @@ DEF_OP(VFMax) {
         mov(GetDst(Node).V2D(), VTMP2.V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -669,7 +669,7 @@ DEF_OP(VFRecp) {
         fdiv(GetDst(Node).D(), VTMP1.D(), GetSrc(Op->Header.Args[0].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -685,7 +685,7 @@ DEF_OP(VFRecp) {
         fdiv(GetDst(Node).V2D(), VTMP1.V2D(), GetSrc(Op->Header.Args[0].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -704,7 +704,7 @@ DEF_OP(VFSqrt) {
         fsqrt(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -718,7 +718,7 @@ DEF_OP(VFSqrt) {
         fsqrt(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -741,7 +741,7 @@ DEF_OP(VFRSqrt) {
         fdiv(GetDst(Node).D(), VTMP1.D(), VTMP2.D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -759,7 +759,7 @@ DEF_OP(VFRSqrt) {
         fdiv(GetDst(Node).V2D(), VTMP1.V2D(), VTMP2.V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -779,7 +779,7 @@ DEF_OP(VNeg) {
   case 8:
     neg(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D());
     break;
-  default: LOGMAN_MSG_A("Unsupported Not size: %d", IROp->Size);
+  default: LOGMAN_MSG_A_FMT("Unsupported VNeg size: {}", IROp->Size);
   }
 }
 
@@ -792,7 +792,7 @@ DEF_OP(VFNeg) {
   case 8:
     fneg(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D());
     break;
-  default: LOGMAN_MSG_A("Unsupported Not size: %d", IROp->Size);
+  default: LOGMAN_MSG_A_FMT("Unsupported VFNeg size: {}", IROp->Size);
   }
 }
 
@@ -823,7 +823,7 @@ DEF_OP(VUMin) {
       mov(GetDst(Node).V2D(), VTMP2.V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -849,7 +849,7 @@ DEF_OP(VSMin) {
       mov(GetDst(Node).V2D(), VTMP2.V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -875,7 +875,7 @@ DEF_OP(VUMax) {
       mov(GetDst(Node).V2D(), VTMP2.V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -901,7 +901,7 @@ DEF_OP(VSMax) {
       mov(GetDst(Node).V2D(), VTMP2.V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -922,7 +922,7 @@ DEF_OP(VZip) {
         zip1(GetDst(Node).V2S(), GetSrc(Op->Header.Args[0].ID()).V2S(), GetSrc(Op->Header.Args[1].ID()).V2S());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -943,7 +943,7 @@ DEF_OP(VZip) {
         zip1(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -965,7 +965,7 @@ DEF_OP(VZip2) {
       zip2(GetDst(Node).V2S(), GetSrc(Op->Header.Args[0].ID()).V2S(), GetSrc(Op->Header.Args[1].ID()).V2S());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -986,7 +986,7 @@ DEF_OP(VZip2) {
       zip2(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1008,7 +1008,7 @@ DEF_OP(VUnZip) {
         uzp1(GetDst(Node).V2S(), GetSrc(Op->Header.Args[0].ID()).V2S(), GetSrc(Op->Header.Args[1].ID()).V2S());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1029,7 +1029,7 @@ DEF_OP(VUnZip) {
         uzp1(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1051,7 +1051,7 @@ DEF_OP(VUnZip2) {
       uzp2(GetDst(Node).V2S(), GetSrc(Op->Header.Args[0].ID()).V2S(), GetSrc(Op->Header.Args[1].ID()).V2S());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1072,7 +1072,7 @@ DEF_OP(VUnZip2) {
       uzp2(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1105,7 +1105,7 @@ DEF_OP(VCMPEQ) {
         cmeq(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), GetSrc(Op->Header.Args[1].ID()).D());
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1127,7 +1127,7 @@ DEF_OP(VCMPEQ) {
         cmeq(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1146,7 +1146,7 @@ DEF_OP(VCMPEQZ) {
         cmeq(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), 0);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1168,7 +1168,7 @@ DEF_OP(VCMPEQZ) {
         cmeq(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), 0);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1187,7 +1187,7 @@ DEF_OP(VCMPGT) {
         cmgt(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), GetSrc(Op->Header.Args[1].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1209,7 +1209,7 @@ DEF_OP(VCMPGT) {
         cmgt(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1228,7 +1228,7 @@ DEF_OP(VCMPGTZ) {
         cmgt(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), 0);
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1250,7 +1250,7 @@ DEF_OP(VCMPGTZ) {
         cmgt(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), 0);
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1269,7 +1269,7 @@ DEF_OP(VCMPLTZ) {
         cmlt(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), 0);
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1291,7 +1291,7 @@ DEF_OP(VCMPLTZ) {
         cmlt(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), 0);
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1310,7 +1310,7 @@ DEF_OP(VFCMPEQ) {
         fcmeq(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), GetSrc(Op->Header.Args[1].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1328,7 +1328,7 @@ DEF_OP(VFCMPEQ) {
         fcmeq(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1347,7 +1347,7 @@ DEF_OP(VFCMPNEQ) {
         fcmeq(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), GetSrc(Op->Header.Args[1].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
     mvn(GetDst(Node).V8B(), GetDst(Node).V8B());
   }
@@ -1366,7 +1366,7 @@ DEF_OP(VFCMPNEQ) {
         fcmeq(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
     mvn(GetDst(Node).V16B(), GetDst(Node).V16B());
   }
@@ -1386,7 +1386,7 @@ DEF_OP(VFCMPLT) {
         fcmgt(GetDst(Node).D(), GetSrc(Op->Header.Args[1].ID()).D(), GetSrc(Op->Header.Args[0].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1404,7 +1404,7 @@ DEF_OP(VFCMPLT) {
         fcmgt(GetDst(Node).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1423,7 +1423,7 @@ DEF_OP(VFCMPGT) {
         fcmgt(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).D(), GetSrc(Op->Header.Args[1].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1441,7 +1441,7 @@ DEF_OP(VFCMPGT) {
         fcmgt(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1460,7 +1460,7 @@ DEF_OP(VFCMPLE) {
         fcmge(GetDst(Node).D(), GetSrc(Op->Header.Args[1].ID()).D(), GetSrc(Op->Header.Args[0].ID()).D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1478,7 +1478,7 @@ DEF_OP(VFCMPLE) {
         fcmge(GetDst(Node).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1501,7 +1501,7 @@ DEF_OP(VFCMPORD) {
         orr(GetDst(Node).V8B(), VTMP1.V8B(), VTMP2.V8B());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1525,7 +1525,7 @@ DEF_OP(VFCMPORD) {
         orr(GetDst(Node).V16B(), VTMP1.V16B(), VTMP2.V16B());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1550,7 +1550,7 @@ DEF_OP(VFCMPUNO) {
         mvn(GetDst(Node).V8B(), GetDst(Node).V8B());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -1577,21 +1577,21 @@ DEF_OP(VFCMPUNO) {
         mvn(GetDst(Node).V16B(), GetDst(Node).V16B());
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
 
 DEF_OP(VUShl) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(VUShr) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(VSShr) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(VUShlS) {
@@ -1618,7 +1618,7 @@ DEF_OP(VUShlS) {
       ushl(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), VTMP1.V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1650,7 +1650,7 @@ DEF_OP(VUShrS) {
       ushl(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), VTMP1.V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1682,7 +1682,7 @@ DEF_OP(VSShrS) {
       sshl(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), VTMP1.V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1713,7 +1713,7 @@ DEF_OP(VInsElement) {
       mov(reg.V2D(), Op->DestIdx, GetSrc(Op->Header.Args[1].ID()).V2D(), Op->SrcIdx);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 
   if (GetDst(Node).GetCode() != reg.GetCode()) {
@@ -1748,7 +1748,7 @@ DEF_OP(VInsScalarElement) {
       mov(reg.V2D(), Op->DestIdx, GetSrc(Op->Header.Args[1].ID()).V2D(), 0);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 
   if (GetDst(Node).GetCode() != reg.GetCode()) {
@@ -1772,7 +1772,7 @@ DEF_OP(VExtractElement) {
     case 8:
       mov(GetDst(Node).D(), GetSrc(Op->Header.Args[0].ID()).V2D(), Op->Index);
     break;
-    default:  LOGMAN_MSG_A("Unhandled ExtractElementSize: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled VExtractElement element size: {}", OpSize);
   }
 }
 
@@ -1791,7 +1791,7 @@ DEF_OP(VDupElement) {
     case 8:
       dup(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), Op->Index);
     break;
-    default:  LOGMAN_MSG_A("Unhandled DupElementSize: %d", Op->Header.ElementSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled VDupElement element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -1906,7 +1906,7 @@ DEF_OP(VUShrI) {
         ushr(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), Op->BitShift);
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1931,7 +1931,7 @@ DEF_OP(VSShrI) {
       sshr(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), std::min((uint8_t)(Op->Header.ElementSize * 8 - 1), Op->BitShift));
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1959,7 +1959,7 @@ DEF_OP(VShlI) {
         shl(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), Op->BitShift);
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1980,7 +1980,7 @@ DEF_OP(VUShrNI) {
       shrn(GetDst(Node).V2S(), GetSrc(Op->Header.Args[0].ID()).V2D(), Op->BitShift);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -2000,7 +2000,7 @@ DEF_OP(VUShrNI2) {
       shrn2(VTMP1.V4S(), GetSrc(Op->Header.Args[1].ID()).V2D(), Op->BitShift);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 
   mov(GetDst(Node), VTMP1);
@@ -2023,7 +2023,7 @@ DEF_OP(VSXTL) {
     case 8:
       sxtl(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2S());
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -2039,7 +2039,7 @@ DEF_OP(VSXTL2) {
     case 8:
       sxtl2(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V4S());
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -2055,7 +2055,7 @@ DEF_OP(VUXTL) {
     case 8:
       uxtl(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2S());
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -2071,7 +2071,7 @@ DEF_OP(VUXTL2) {
     case 8:
       uxtl2(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V4S());
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -2087,7 +2087,7 @@ DEF_OP(VSQXTN) {
     case 4:
       sqxtn(GetDst(Node).V2S(), GetSrc(Op->Header.Args[0].ID()).V2D());
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -2109,7 +2109,7 @@ DEF_OP(VSQXTN2) {
         sqxtn(VTMP2.V2S(), GetSrc(Op->Header.Args[1].ID()).V2D());
         ins(VTMP1.V4S(), 1, VTMP2.V4S(), 0);
       break;
-      default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
     }
   }
   else {
@@ -2123,7 +2123,7 @@ DEF_OP(VSQXTN2) {
       case 4:
         sqxtn2(VTMP1.V4S(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
-      default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
     }
   }
   mov(GetDst(Node), VTMP1);
@@ -2141,7 +2141,7 @@ DEF_OP(VSQXTUN) {
     case 4:
       sqxtun(GetDst(Node).V2S(), GetSrc(Op->Header.Args[0].ID()).V2D());
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -2163,7 +2163,7 @@ DEF_OP(VSQXTUN2) {
         sqxtun(VTMP2.V2S(), GetSrc(Op->Header.Args[1].ID()).V2D());
         ins(VTMP1.V4S(), 1, VTMP2.V4S(), 0);
       break;
-      default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
     }
   }
   else {
@@ -2177,7 +2177,7 @@ DEF_OP(VSQXTUN2) {
       case 4:
         sqxtun2(VTMP1.V4S(), GetSrc(Op->Header.Args[1].ID()).V2D());
       break;
-      default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
     }
   }
   mov(GetDst(Node), VTMP1);
@@ -2202,7 +2202,7 @@ DEF_OP(VMul) {
       mul(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2D(), GetSrc(Op->Header.Args[1].ID()).V2D());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -2221,7 +2221,7 @@ DEF_OP(VUMull) {
       umull(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2S(), GetSrc(Op->Header.Args[1].ID()).V2S());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize >> 1); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize >> 1); break;
   }
 }
 
@@ -2240,7 +2240,7 @@ DEF_OP(VSMull) {
       smull(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2S(), GetSrc(Op->Header.Args[1].ID()).V2S());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize >> 1); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize >> 1); break;
   }
 }
 
@@ -2259,7 +2259,7 @@ DEF_OP(VUMull2) {
       umull2(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V4S(), GetSrc(Op->Header.Args[1].ID()).V4S());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize >> 1); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize >> 1); break;
   }
 }
 
@@ -2278,7 +2278,7 @@ DEF_OP(VSMull2) {
       smull2(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V4S(), GetSrc(Op->Header.Args[1].ID()).V4S());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize >> 1); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize >> 1); break;
   }
 }
 
@@ -2297,7 +2297,7 @@ DEF_OP(VUABDL) {
       uabdl(GetDst(Node).V2D(), GetSrc(Op->Header.Args[0].ID()).V2S(), GetSrc(Op->Header.Args[1].ID()).V2S());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize >> 1); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize >> 1); break;
   }
 }
 
@@ -2314,7 +2314,7 @@ DEF_OP(VTBL1) {
       tbl(GetDst(Node).V16B(), GetSrc(Op->Header.Args[0].ID()).V16B(), GetSrc(Op->Header.Args[1].ID()).V16B());
     break;
     }
-    default: LOGMAN_MSG_A("Unknown OpSize: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown OpSize: {}", OpSize); break;
   }
 }
 

External/FEXCore/Source/Interface/Core/JIT/x86_64/ALUOps.cpp

@@ -5,7 +5,14 @@ $end_info$
 */
 
 #include "Interface/Core/JIT/x86_64/JITClass.h"
-#include "Interface/IR/Passes/RegisterAllocationPass.h"
+
+#include <FEXCore/IR/IR.h>
+#include <FEXCore/Utils/LogManager.h>
+
+#include <array>
+#include <stdint.h>
+#include <utility>
+#include <xbyak/xbyak.h>
 
 namespace FEXCore::CPU {
 #define DEF_OP(x) void X86JITCore::Op_##x(FEXCore::IR::IROp_Header *IROp, uint32_t Node)
@@ -20,7 +27,7 @@ DEF_OP(TruncElementPair) {
       mov(Dst.second, Src.second);
       break;
     }
-    default: LOGMAN_MSG_A("Unhandled Truncation size: %d", Op->Size); break;
+    default: LOGMAN_MSG_A_FMT("Unhandled Truncation size: {}", Op->Size); break;
   }
 }
 
@@ -70,7 +77,7 @@ DEF_OP(Add) {
     case 8:
       add(rax, Const);
       break;
-    default:  LOGMAN_MSG_A("Unhandled Add size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Add size: {}", OpSize);
       break;
     }
   } else {
@@ -81,7 +88,7 @@ DEF_OP(Add) {
     case 8:
       add(rax, GetSrc<RA_64>(Op->Header.Args[1].ID()));
       break;
-    default:  LOGMAN_MSG_A("Unhandled Add size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Add size: {}", OpSize);
       break;
     }
   }
@@ -103,7 +110,7 @@ DEF_OP(Sub) {
     case 8:
       sub(rax, Const);
       break;
-    default:  LOGMAN_MSG_A("Unhandled Sub size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Sub size: {}", OpSize);
       break;
     }
   } else {
@@ -114,7 +121,7 @@ DEF_OP(Sub) {
     case 8:
       sub(rax, GetSrc<RA_64>(Op->Header.Args[1].ID()));
       break;
-    default:  LOGMAN_MSG_A("Unhandled Sub size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled Sub size: {}", OpSize);
       break;
     }
   }
@@ -136,7 +143,7 @@ DEF_OP(Neg) {
     Src = GetSrc<RA_64>(Op->Header.Args[0].ID());
     Dst = GetDst<RA_64>(Node);
     break;
-  default:  LOGMAN_MSG_A("Unhandled Neg size: %d", OpSize);
+  default:  LOGMAN_MSG_A_FMT("Unhandled Neg size: {}", OpSize);
     break;
   }
   mov(Dst, Src);
@@ -160,7 +167,7 @@ DEF_OP(Mul) {
     imul(rax, GetSrc<RA_64>(Op->Header.Args[1].ID()));
     mov(Dst, rax);
   break;
-  default: LOGMAN_MSG_A("Unknown Sext size: %d", OpSize);
+  default: LOGMAN_MSG_A_FMT("Unknown Mul size: {}", OpSize);
   }
 }
 
@@ -179,7 +186,7 @@ DEF_OP(UMul) {
     mul(GetSrc<RA_64>(Op->Header.Args[1].ID()));
     mov(GetDst<RA_64>(Node), rax);
   break;
-  default: LOGMAN_MSG_A("Unknown Sext size: %d", OpSize);
+  default: LOGMAN_MSG_A_FMT("Unknown UMul size: {}", OpSize);
   }
 }
 
@@ -218,7 +225,7 @@ DEF_OP(Div) {
     mov(GetDst<RA_64>(Node), rax);
   break;
   }
-  default: LOGMAN_MSG_A("Unknown UDIV Size: %d", Size); break;
+  default: LOGMAN_MSG_A_FMT("Unknown DIV Size: {}", Size); break;
   }
 }
 
@@ -261,7 +268,7 @@ DEF_OP(UDiv) {
     mov(GetDst<RA_64>(Node), rax);
   break;
   }
-  default: LOGMAN_MSG_A("Unknown UDIV OpSize: %d", OpSize); break;
+  default: LOGMAN_MSG_A_FMT("Unknown UDIV OpSize: {}", OpSize); break;
   }
 }
 
@@ -298,7 +305,7 @@ DEF_OP(Rem) {
     mov(GetDst<RA_64>(Node), rdx);
   break;
   }
-  default: LOGMAN_MSG_A("Unknown UDIV Size: %d", OpSize); break;
+  default: LOGMAN_MSG_A_FMT("Unknown Rem Size: {}", OpSize); break;
   }
 }
 
@@ -341,7 +348,7 @@ DEF_OP(URem) {
     mov(GetDst<RA_64>(Node), rdx);
   break;
   }
-  default: LOGMAN_MSG_A("Unknown UDIV OpSize: %d", OpSize); break;
+  default: LOGMAN_MSG_A_FMT("Unknown URem OpSize: {}", OpSize); break;
   }
 }
 
@@ -360,7 +367,7 @@ DEF_OP(MulH) {
     imul(GetSrc<RA_64>(Op->Header.Args[1].ID()));
     mov(GetDst<RA_64>(Node), rdx);
   break;
-  default: LOGMAN_MSG_A("Unknown Sext size: %d", OpSize);
+  default: LOGMAN_MSG_A_FMT("Unknown MulH size: {}", OpSize);
   }
 }
 
@@ -379,7 +386,7 @@ DEF_OP(UMulH) {
     mul(GetSrc<RA_64>(Op->Header.Args[1].ID()));
     mov(GetDst<RA_64>(Node), rdx);
   break;
-  default: LOGMAN_MSG_A("Unknown Sext size: %d", OpSize);
+  default: LOGMAN_MSG_A_FMT("Unknown UMulH size: {}", OpSize);
   }
 }
 
@@ -441,7 +448,7 @@ DEF_OP(Lshl) {
         mov(GetDst<RA_64>(Node), GetSrc<RA_64>(Op->Header.Args[0].ID()));
         shl(GetDst<RA_64>(Node), Const);
         break;
-      default: LOGMAN_MSG_A("Unknown LSHL Size: %d\n", OpSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown LSHL Size: {}\n", OpSize); break;
     };
   } else {
     mov(rcx, GetSrc<RA_64>(Op->Header.Args[1].ID()));
@@ -456,7 +463,7 @@ DEF_OP(Lshl) {
         mov(GetDst<RA_64>(Node), GetSrc<RA_64>(Op->Header.Args[0].ID()));
         shl(GetDst<RA_64>(Node), cl);
         break;
-      default: LOGMAN_MSG_A("Unknown LSHL Size: %d\n", OpSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown LSHL Size: {}\n", OpSize); break;
     };
   }
 }
@@ -488,7 +495,7 @@ DEF_OP(Lshr) {
         mov(GetDst<RA_64>(Node), GetSrc<RA_64>(Op->Header.Args[0].ID()));
         shr(GetDst<RA_64>(Node), Const);
         break;
-      default: LOGMAN_MSG_A("Unknown Size: %d\n", OpSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Size: {}\n", OpSize); break;
     };
 
   } else {
@@ -512,7 +519,7 @@ DEF_OP(Lshr) {
         mov(GetDst<RA_64>(Node), GetSrc<RA_64>(Op->Header.Args[0].ID()));
         shr(GetDst<RA_64>(Node), cl);
         break;
-      default: LOGMAN_MSG_A("Unknown Size: %d\n", OpSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Size: {}\n", OpSize); break;
     };
   }
 }
@@ -546,7 +553,7 @@ DEF_OP(Ashr) {
       mov(GetDst<RA_64>(Node), GetSrc<RA_64>(Op->Header.Args[0].ID()));
       sar(GetDst<RA_64>(Node), Const);
     break;
-    default: LOGMAN_MSG_A("Unknown ASHR Size: %d\n", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown ASHR Size: {}\n", OpSize); break;
     };
 
   } else {
@@ -571,7 +578,7 @@ DEF_OP(Ashr) {
       mov(GetDst<RA_64>(Node), GetSrc<RA_64>(Op->Header.Args[0].ID()));
       sar(GetDst<RA_64>(Node), cl);
     break;
-    default: LOGMAN_MSG_A("Unknown ASHR Size: %d\n", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown ASHR Size: {}\n", OpSize); break;
     };
   }
 }
@@ -596,7 +603,7 @@ DEF_OP(Ror) {
         ror(rax, Const);
       break;
       }
-      default: LOGMAN_MSG_A("Unknown ROR Size: %d\n", OpSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown ROR Size: {}\n", OpSize); break;
     }
   } else {
     mov (rcx, GetSrc<RA_64>(Op->Header.Args[1].ID()));
@@ -612,7 +619,7 @@ DEF_OP(Ror) {
         ror(rax, cl);
       break;
       }
-      default: LOGMAN_MSG_A("Unknown ROR Size: %d\n", OpSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown ROR Size: {}\n", OpSize); break;
     }
   }
   mov(GetDst<RA_64>(Node), rax);
@@ -668,7 +675,7 @@ DEF_OP(LDiv) {
       mov(GetDst<RA_64>(Node), rax);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown LDIV OpSize: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown LDIV OpSize: {}", OpSize); break;
   }
 }
 
@@ -700,7 +707,7 @@ DEF_OP(LUDiv) {
       mov(GetDst<RA_64>(Node), rax);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown LUDIV OpSize: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown LUDIV OpSize: {}", OpSize); break;
   }
 }
 
@@ -732,7 +739,7 @@ DEF_OP(LRem) {
       mov(GetDst<RA_64>(Node), rdx);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown LREM OpSize: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown LREM OpSize: {}", OpSize); break;
   }
 }
 
@@ -764,7 +771,7 @@ DEF_OP(LURem) {
       mov(GetDst<RA_64>(Node), rdx);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown LUDIV OpSize: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown LUREM OpSize: {}", OpSize); break;
   }
 }
 
@@ -829,7 +836,7 @@ DEF_OP(FindMSB) {
     case 8:
       bsr(GetDst<RA_64>(Node), GetSrc<RA_64>(Op->Header.Args[0].ID()));
       break;
-    default: LOGMAN_MSG_A("Unknown OpSize: %d", OpSize);
+    default: LOGMAN_MSG_A_FMT("Unknown FindMSB OpSize: {}", OpSize);
   }
 }
 
@@ -853,7 +860,7 @@ DEF_OP(FindTrailingZeros) {
       mov(rax, 0x40);
       cmovz(GetDst<RA_64>(Node), rax);
       break;
-    default: LOGMAN_MSG_A("Unknown size: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown FindTrailingZeros size: {}", OpSize); break;
   }
 }
 
@@ -876,7 +883,7 @@ DEF_OP(CountLeadingZeroes) {
         lzcnt(GetDst<RA_64>(Node), GetSrc<RA_64>(Op->Header.Args[0].ID()));
         break;
       }
-      default: LOGMAN_MSG_A("Unknown size: %d", OpSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown CountLeadingZeros size: {}", OpSize); break;
     }
   }
   else {
@@ -915,7 +922,7 @@ DEF_OP(CountLeadingZeroes) {
         mov(GetDst<RA_64>(Node), rax);
         break;
       }
-      default: LOGMAN_MSG_A("Unknown size: %d", OpSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown CountLeadingZeros size: {}", OpSize); break;
     }
   }
 }
@@ -937,7 +944,7 @@ DEF_OP(Rev) {
       mov (GetDst<RA_64>(Node), GetSrc<RA_64>(Op->Header.Args[0].ID()));
       bswap(GetDst<RA_64>(Node).cvt64());
       break;
-    default: LOGMAN_MSG_A("Unknown REV size: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown REV size: {}", OpSize); break;
   }
 }
 
@@ -970,7 +977,7 @@ DEF_OP(Bfi) {
 
 DEF_OP(Bfe) {
   auto Op = IROp->C<IR::IROp_Bfe>();
-  LOGMAN_THROW_A(IROp->Size <= 8, "OpSize is too large for BFE: %d", IROp->Size);
+  LOGMAN_THROW_A_FMT(IROp->Size <= 8, "OpSize is too large for BFE: {}", IROp->Size);
 
   auto Dst = GetDst<RA_64>(Node);
 
@@ -1071,7 +1078,7 @@ DEF_OP(Select) {
 
   if (is_const_true || is_const_false) {
     if (is_const_false != true || is_const_true != true || const_true != 1 || const_false != 0) {
-      LOGMAN_MSG_A("Select: Unsupported compare inline parameters");
+      LOGMAN_MSG_A_FMT("Select: Unsupported compare inline parameters");
     }
     (this->*SetCC)(al);
     movzx(Dst, al);
@@ -1102,7 +1109,7 @@ DEF_OP(VExtractToGPR) {
       pextrq(GetDst<RA_64>(Node), GetSrc(Op->Header.Args[0].ID()), Op->Idx);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 

External/FEXCore/Source/Interface/Core/JIT/x86_64/AtomicOps.cpp

@@ -5,7 +5,14 @@ $end_info$
 */
 
 #include "Interface/Core/JIT/x86_64/JITClass.h"
-#include "Interface/IR/Passes/RegisterAllocationPass.h"
+
+#include <FEXCore/IR/IR.h>
+#include <FEXCore/Utils/LogManager.h>
+
+#include <array>
+#include <stdint.h>
+#include <utility>
+#include <xbyak/xbyak.h>
 
 namespace FEXCore::CPU {
 #define DEF_OP(x) void X86JITCore::Op_##x(FEXCore::IR::IROp_Header *IROp, uint32_t Node)
@@ -55,7 +62,7 @@ DEF_OP(CASPair) {
       mov(Dst.second, rdx);
     break;
     }
-    default: LOGMAN_MSG_A("Unsupported: %d", OpSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported: {}", OpSize);
   }
 }
 
@@ -104,7 +111,7 @@ DEF_OP(CAS) {
     mov (GetDst<RA_64>(Node), rax);
   break;
   }
-  default: LOGMAN_MSG_A("Unsupported: %d", OpSize);
+  default: LOGMAN_MSG_A_FMT("Unsupported: {}", OpSize);
   }
 }
 
@@ -127,7 +134,7 @@ DEF_OP(AtomicAdd) {
     case 8:
       add(qword [MemReg], GetSrc<RA_64>(Op->Header.Args[1].ID()));
       break;
-    default:  LOGMAN_MSG_A("Unhandled AtomicAdd size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled AtomicAdd size: {}", Op->Size);
   }
 }
 
@@ -149,7 +156,7 @@ DEF_OP(AtomicSub) {
     case 8:
       sub(qword [MemReg], GetSrc<RA_64>(Op->Header.Args[1].ID()));
       break;
-    default:  LOGMAN_MSG_A("Unhandled AtomicAdd size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled AtomicAdd size: {}", Op->Size);
   }
 }
 
@@ -171,7 +178,7 @@ DEF_OP(AtomicAnd) {
     case 8:
       and_(qword [MemReg], GetSrc<RA_64>(Op->Header.Args[1].ID()));
       break;
-    default:  LOGMAN_MSG_A("Unhandled AtomicAdd size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled AtomicAdd size: {}", Op->Size);
   }
 }
 
@@ -193,7 +200,7 @@ DEF_OP(AtomicOr) {
     case 8:
       or_(qword [MemReg], GetSrc<RA_64>(Op->Header.Args[1].ID()));
       break;
-    default:  LOGMAN_MSG_A("Unhandled AtomicAdd size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled AtomicAdd size: {}", Op->Size);
   }
 }
 
@@ -215,7 +222,7 @@ DEF_OP(AtomicXor) {
     case 8:
       xor_(qword [MemReg], GetSrc<RA_64>(Op->Header.Args[1].ID()));
       break;
-    default:  LOGMAN_MSG_A("Unhandled AtomicAdd size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled AtomicAdd size: {}", Op->Size);
   }
 }
 
@@ -246,7 +253,7 @@ DEF_OP(AtomicSwap) {
       lock();
       xchg(qword [MemReg], GetDst<RA_64>(Node));
       break;
-    default:  LOGMAN_MSG_A("Unhandled AtomicAdd size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled AtomicSwap size: {}", Op->Size);
   }
 }
 
@@ -279,7 +286,7 @@ DEF_OP(AtomicFetchAdd) {
       xadd(qword [MemReg], rcx);
       mov(GetDst<RA_64>(Node), rcx);
       break;
-    default:  LOGMAN_MSG_A("Unhandled AtomicFetchAdd size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled AtomicFetchAdd size: {}", Op->Size);
   }
 }
 
@@ -316,7 +323,7 @@ DEF_OP(AtomicFetchSub) {
       xadd(qword [MemReg], rcx);
       mov(GetDst<RA_64>(Node), rcx);
       break;
-    default:  LOGMAN_MSG_A("Unhandled AtomicFetchAdd size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled AtomicFetchSub size: {}", Op->Size);
   }
 }
 
@@ -394,7 +401,7 @@ DEF_OP(AtomicFetchAnd) {
       mov(GetDst<RA_64>(Node), TMP3.cvt64());
       break;
     }
-    default:  LOGMAN_MSG_A("Unhandled AtomicFetchAdd size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled AtomicFetchAnd size: {}", Op->Size);
   }
 }
 
@@ -471,7 +478,7 @@ DEF_OP(AtomicFetchOr) {
       mov(GetDst<RA_64>(Node), TMP3.cvt64());
       break;
     }
-    default:  LOGMAN_MSG_A("Unhandled AtomicFetchAdd size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled AtomicFetchOr size: {}", Op->Size);
   }
 }
 
@@ -548,7 +555,83 @@ DEF_OP(AtomicFetchXor) {
       mov(GetDst<RA_64>(Node), TMP3.cvt64());
       break;
     }
-    default:  LOGMAN_MSG_A("Unhandled AtomicFetchAdd size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled AtomicFetchXor size: {}", Op->Size);
+  }
+}
+
+DEF_OP(AtomicFetchNeg) {
+  auto Op = IROp->C<IR::IROp_AtomicFetchNeg>();
+
+  Xbyak::Reg MemReg = GetSrc<RA_64>(Op->Header.Args[0].ID());
+  switch (Op->Size) {
+    case 1: {
+      mov(TMP1.cvt8(), byte [MemReg]);
+
+      Label Loop;
+      L(Loop);
+      mov(TMP2.cvt8(), TMP1.cvt8());
+      mov(TMP3.cvt8(), TMP1.cvt8());
+      neg(TMP2.cvt8());
+
+      // Updates RAX with the value from memory
+      lock(); cmpxchg(byte [MemReg], TMP2.cvt8());
+      jne(Loop);
+      // Result is the previous value from memory, which is currently in TMP3
+      movzx(GetDst<RA_64>(Node), TMP3.cvt8());
+      break;
+    }
+    case 2: {
+      mov(TMP1.cvt16(), word [MemReg]);
+
+      Label Loop;
+      L(Loop);
+      mov(TMP2.cvt16(), TMP1.cvt16());
+      mov(TMP3.cvt16(), TMP1.cvt16());
+      neg(TMP2.cvt16());
+
+      // Updates RAX with the value from memory
+      lock(); cmpxchg(word [MemReg], TMP2.cvt16());
+      jne(Loop);
+
+      // Result is the previous value from memory, which is currently in TMP3
+      movzx(GetDst<RA_64>(Node), TMP3.cvt16());
+      break;
+    }
+    case 4: {
+      mov(TMP1.cvt32(), dword [MemReg]);
+
+      Label Loop;
+      L(Loop);
+      mov(TMP2.cvt32(), TMP1.cvt32());
+      mov(TMP3.cvt32(), TMP1.cvt32());
+      neg(TMP2.cvt32());
+
+      // Updates RAX with the value from memory
+      lock(); cmpxchg(dword [MemReg], TMP2.cvt32());
+      jne(Loop);
+
+      // Result is the previous value from memory, which is currently in TMP3
+      mov(GetDst<RA_32>(Node), TMP3.cvt32());
+      break;
+    }
+    case 8: {
+      mov(TMP1.cvt64(), qword [MemReg]);
+
+      Label Loop;
+      L(Loop);
+      mov(TMP2.cvt64(), TMP1.cvt64());
+      mov(TMP3.cvt64(), TMP1.cvt64());
+      neg(TMP2.cvt64());
+
+      // Updates RAX with the value from memory
+      lock(); cmpxchg(qword [MemReg], TMP2.cvt64());
+      jne(Loop);
+
+      // Result is the previous value from memory, which is currently in TMP3
+      mov(GetDst<RA_64>(Node), TMP3.cvt64());
+      break;
+    }
+    default:  LOGMAN_MSG_A_FMT("Unhandled AtomicFetchNeg size: {}", Op->Size);
   }
 }
 
@@ -568,6 +651,7 @@ void X86JITCore::RegisterAtomicHandlers() {
   REGISTER_OP(ATOMICFETCHAND, AtomicFetchAnd);
   REGISTER_OP(ATOMICFETCHOR,  AtomicFetchOr);
   REGISTER_OP(ATOMICFETCHXOR, AtomicFetchXor);
+  REGISTER_OP(ATOMICFETCHNEG, AtomicFetchNeg);
 #undef REGISTER_OP
 }
 }

External/FEXCore/Source/Interface/Core/JIT/x86_64/BranchOps.cpp

@@ -4,25 +4,41 @@ tags: backend|x86-64
 $end_info$
 */
 
+#include "Interface/Context/Context.h"
+#include "Interface/Core/CPUID.h"
+#include "Interface/Core/Dispatcher/Dispatcher.h"
+#include "Interface/Core/LookupCache.h"
 #include "Interface/Core/JIT/x86_64/JITClass.h"
-#include "Interface/IR/Passes/RegisterAllocationPass.h"
+#include "Interface/HLE/Thunks/Thunks.h"
 
+#include <FEXCore/Core/CPUID.h>
+#include <FEXCore/Core/CoreState.h>
 #include <FEXCore/Core/X86Enums.h>
+#include <FEXCore/Debug/InternalThreadState.h>
 #include <FEXCore/HLE/SyscallHandler.h>
-#include <Interface/HLE/Thunks/Thunks.h>
+#include <FEXCore/IR/IR.h>
+#include <FEXCore/Utils/LogManager.h>
+
+#include <array>
+#include <memory>
+#include <stddef.h>
+#include <stdint.h>
+#include <unordered_map>
+#include <utility>
+#include <xbyak/xbyak.h>
 
 namespace FEXCore::CPU {
 #define DEF_OP(x) void X86JITCore::Op_##x(FEXCore::IR::IROp_Header *IROp, uint32_t Node)
 DEF_OP(GuestCallDirect) {
-  LogMan::Msg::D("Unimplemented");
+  LogMan::Msg::DFmt("Unimplemented");
 }
 
 DEF_OP(GuestCallIndirect) {
-  LogMan::Msg::D("Unimplemented");
+  LogMan::Msg::DFmt("Unimplemented");
 }
 
 DEF_OP(GuestReturn) {
-  LogMan::Msg::D("Unimplemented");
+  LogMan::Msg::DFmt("Unimplemented");
 }
 
 DEF_OP(SignalReturn) {

External/FEXCore/Source/Interface/Core/JIT/x86_64/ConversionOps.cpp

@@ -5,7 +5,13 @@ $end_info$
 */
 
 #include "Interface/Core/JIT/x86_64/JITClass.h"
-#include "Interface/IR/Passes/RegisterAllocationPass.h"
+
+#include <FEXCore/IR/IR.h>
+#include <FEXCore/Utils/LogManager.h>
+
+#include <array>
+#include <stdint.h>
+#include <xbyak/xbyak.h>
 
 namespace FEXCore::CPU {
 
@@ -31,7 +37,7 @@ DEF_OP(VInsGPR) {
       pinsrq(GetDst(Node), GetSrc<RA_64>(Op->Header.Args[1].ID()), Op->Index);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -52,7 +58,7 @@ DEF_OP(VCastFromGPR) {
     case 8:
       vmovq(GetDst(Node), GetSrc<RA_64>(Op->Header.Args[0].ID()).cvt64());
       break;
-    default: LOGMAN_MSG_A("Unknown castGPR element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown VCastFromGPR element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -91,7 +97,7 @@ DEF_OP(Float_FToF) {
       cvtsd2ss(GetDst(Node), GetSrc(Op->Header.Args[0].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown FCVT sizes: 0x%x", Conv);
+    default: LOGMAN_MSG_A_FMT("Unknown Float_FToF sizes: 0x{:x}", Conv);
   }
 }
 
@@ -113,7 +119,7 @@ DEF_OP(Vector_SToF) {
       cvtsi2sd(xmm15, rax);
       movlhps(GetDst(Node), xmm15);
     break;
-    default: LOGMAN_MSG_A("Unknown castGPR element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Vector_SToF element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -126,7 +132,7 @@ DEF_OP(Vector_FToZS) {
     case 8:
       cvttpd2dq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown castGPR element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Vector_FToZS element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -139,7 +145,7 @@ DEF_OP(Vector_FToS) {
     case 8:
       cvtpd2dq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown castGPR element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Vector_FToS element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -156,7 +162,7 @@ DEF_OP(Vector_FToF) {
       cvtpd2ps(GetDst(Node), GetSrc(Op->Header.Args[0].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Conversion Type : 0%04x", Conv); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Vector_FToF conversion type : 0x{:04x}", Conv); break;
   }
 }
 

External/FEXCore/Source/Interface/Core/JIT/x86_64/EncryptionOps.cpp

@@ -5,7 +5,12 @@ $end_info$
 */
 
 #include "Interface/Core/JIT/x86_64/JITClass.h"
-#include "Interface/IR/Passes/RegisterAllocationPass.h"
+
+#include <FEXCore/IR/IR.h>
+
+#include <array>
+#include <stdint.h>
+#include <xbyak/xbyak.h>
 
 namespace FEXCore::CPU {
 #define DEF_OP(x) void X86JITCore::Op_##x(FEXCore::IR::IROp_Header *IROp, uint32_t Node)

External/FEXCore/Source/Interface/Core/JIT/x86_64/FlagOps.cpp

@@ -5,7 +5,12 @@ $end_info$
 */
 
 #include "Interface/Core/JIT/x86_64/JITClass.h"
-#include "Interface/IR/Passes/RegisterAllocationPass.h"
+
+#include <FEXCore/IR/IR.h>
+
+#include <array>
+#include <stdint.h>
+#include <xbyak/xbyak.h>
 
 namespace FEXCore::CPU {
 

External/FEXCore/Source/Interface/Core/JIT/x86_64/JIT.cpp

@@ -6,21 +6,38 @@ $end_info$
 */
 
 #include "Interface/Context/Context.h"
+#include "Interface/Core/LookupCache.h"
 
+#include "Interface/Core/Dispatcher/Dispatcher.h"
 #include "Interface/Core/Dispatcher/X86Dispatcher.h"
+#include "Interface/Core/Interpreter/InterpreterOps.h"
 #include "Interface/Core/JIT/x86_64/JITClass.h"
-#include "Interface/Core/InternalThreadState.h"
-
+#include "Interface/IR/PassManager.h"
 #include "Interface/IR/Passes/RegisterAllocationPass.h"
 
-#include <FEXCore/Core/X86Enums.h>
-#include <FEXCore/Core/UContext.h>
+#include <FEXCore/Core/CPUBackend.h>
+#include <FEXCore/Core/CoreState.h>
+#include <FEXCore/Core/SignalDelegator.h>
+#include <FEXCore/Debug/InternalThreadState.h>
+#include <FEXCore/IR/IR.h>
+#include <FEXCore/IR/IntrusiveIRList.h>
+#include <FEXCore/IR/RegisterAllocationData.h>
 #include <FEXCore/Utils/Allocator.h>
+#include <FEXCore/Utils/LogManager.h>
 
-#include <cmath>
+#include <algorithm>
+#include <array>
+#include <bits/types/stack_t.h>
+#include <memory>
+#include <stddef.h>
+#include <stdint.h>
 #include <signal.h>
-
-#include "Interface/Core/Interpreter/InterpreterOps.h"
+#include <sys/mman.h>
+#include <tuple>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+#include <xbyak/xbyak.h>
 
 // #define DEBUG_RA 1
 // #define DEBUG_CYCLES
@@ -36,7 +53,7 @@ CodeBuffer AllocateNewCodeBuffer(size_t Size) {
                     PROT_READ | PROT_WRITE | PROT_EXEC,
                     MAP_PRIVATE | MAP_ANONYMOUS,
                     -1, 0));
-  LOGMAN_THROW_A(Buffer.Ptr != reinterpret_cast<uint8_t*>(~0ULL), "Couldn't allocate code buffer");
+  LOGMAN_THROW_A_FMT(Buffer.Ptr != reinterpret_cast<uint8_t*>(~0ULL), "Couldn't allocate code buffer");
   return Buffer;
 }
 
@@ -54,9 +71,9 @@ void X86JITCore::CopyNecessaryDataForCompileThread(CPUBackend *Original) {
 }
 
 void X86JITCore::PushRegs() {
-  for (auto &Xmm : RAXMM_x) {
-    sub(rsp, 16);
-    movaps(ptr[rsp], Xmm);
+  sub(rsp, 16 * RAXMM_x.size());
+  for (size_t i = 0; i < RAXMM_x.size(); ++i) {
+    movaps(ptr[rsp + i * 16], RAXMM_x[i]);
   }
 
   for (auto &Reg : RA64)
@@ -75,18 +92,18 @@ void X86JITCore::PopRegs() {
   for (uint32_t i = RA64.size(); i > 0; --i)
     pop(RA64[i - 1]);
 
-  for (uint32_t i = RAXMM_x.size(); i > 0; --i) {
-    movaps(RAXMM_x[i - 1], ptr[rsp]);
-    add(rsp, 16);
+  for (size_t i = 0; i < RAXMM_x.size(); ++i) {
+    movaps(RAXMM_x[i], ptr[rsp + i * 16]);
   }
+
+  add(rsp, 16 * RAXMM_x.size());
 }
 
 void X86JITCore::Op_Unhandled(FEXCore::IR::IROp_Header *IROp, uint32_t Node) {
   FallbackInfo Info;
   if (!InterpreterOps::GetFallbackHandler(IROp, &Info)) {
 #if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
-    auto Name = FEXCore::IR::GetName(IROp->Op);
-    LOGMAN_MSG_A("Unhandled IR Op: %s", std::string(Name).c_str());
+    LOGMAN_MSG_A_FMT("Unhandled IR Op: {}", FEXCore::IR::GetName(IROp->Op));
 #endif
   } else {
     switch(Info.ABI) {
@@ -285,8 +302,7 @@ void X86JITCore::Op_Unhandled(FEXCore::IR::IROp_Header *IROp, uint32_t Node) {
       case FABI_UNKNOWN:
       default:
 #if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
-      auto Name = FEXCore::IR::GetName(IROp->Op);
-        LOGMAN_MSG_A("Unhandled IR Fallback abi: %s %d", std::string(Name).c_str(), Info.ABI);
+        LOGMAN_MSG_A_FMT("Unhandled IR Fallback ABI: {} {}", FEXCore::IR::GetName(IROp->Op), Info.ABI);
 #endif
         break;
     }
@@ -304,7 +320,7 @@ X86JITCore::X86JITCore(FEXCore::Context::Context *ctx, FEXCore::Core::InternalTh
 {
   CurrentCodeBuffer = &InitialCodeBuffer;
 
-  RAPass = Thread->PassManager->GetRAPass();
+  RAPass = Thread->PassManager->GetPass<IR::RegisterAllocationPass>("RA");
 
   RAPass->AllocateRegisterSet(RegisterCount, RegisterClasses);
   RAPass->AddRegisters(FEXCore::IR::GPRClass, NumGPRs);
@@ -360,7 +376,7 @@ X86JITCore::X86JITCore(FEXCore::Context::Context *ctx, FEXCore::Core::InternalTh
       return Core->Dispatcher->HandleGuestSignal(Signal, info, ucontext, GuestAction, GuestStack);
     };
 
-    for (uint32_t Signal = 0; Signal < SignalDelegator::MAX_SIGNALS; ++Signal) {
+    for (uint32_t Signal = 0; Signal <= SignalDelegator::MAX_SIGNALS; ++Signal) {
       CTX->SignalDelegation->RegisterHostSignalHandlerForGuest(Signal, GuestSignalHandler);
     }
   }
@@ -419,7 +435,7 @@ void X86JITCore::ClearCache() {
 IR::PhysicalRegister X86JITCore::GetPhys(uint32_t Node) const {
   auto PhyReg = RAData->GetNodeRegister(Node);
 
-  LOGMAN_THROW_A(PhyReg.Raw != 255, "Couldn't Allocate register for node: ssa%d. Class: %d", Node, PhyReg.Class);
+  LOGMAN_THROW_A_FMT(PhyReg.Raw != 255, "Couldn't Allocate register for node: ssa{}. Class: {}", Node, PhyReg.Class);
 
   return PhyReg;
 }
@@ -568,7 +584,7 @@ std::tuple<X86JITCore::SetCC, X86JITCore::CMovCC, X86JITCore::JCC> X86JITCore::G
     case FEXCore::IR::COND_VS:
     case FEXCore::IR::COND_VC:
     default:
-      LOGMAN_MSG_A("Unsupported compare type");
+      LOGMAN_MSG_A_FMT("Unsupported compare type");
       break;
   }
 
@@ -611,7 +627,7 @@ void *X86JITCore::CompileCode(uint64_t Entry, [[maybe_unused]] FEXCore::IR::IRLi
     L(RunBlock);
   }
 
-  LOGMAN_THROW_A(RAData != nullptr, "Needs RA");
+  LOGMAN_THROW_A_FMT(RAData != nullptr, "Needs RA");
 
   SpillSlots = RAData->SpillSlots();
 
@@ -669,7 +685,7 @@ void *X86JITCore::CompileCode(uint64_t Entry, [[maybe_unused]] FEXCore::IR::IRLi
     {
 #if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
       auto BlockIROp = BlockHeader->CW<IROp_CodeBlock>();
-      LOGMAN_THROW_A(BlockIROp->Header.Op == IR::OP_CODEBLOCK, "IR type failed to be a code block");
+      LOGMAN_THROW_A_FMT(BlockIROp->Header.Op == IR::OP_CODEBLOCK, "IR type failed to be a code block");
 #endif
 
       uint32_t Node = IR->GetID(BlockNode);
@@ -721,7 +737,7 @@ void *X86JITCore::CompileCode(uint64_t Entry, [[maybe_unused]] FEXCore::IR::IRLi
             Inst << "Reg" << GetPhys(ArgNode) << (i + 1 == NumArgs ? "" : ", ");
         }
 
-        LogMan::Msg::D("%s", Inst.str().c_str());
+        LogMan::Msg::DFmt("{}", Inst.str());
       }
       #endif
       uint32_t ID = IR->GetID(CodeNode);

External/FEXCore/Source/Interface/Core/JIT/x86_64/JITClass.h

@@ -6,7 +6,6 @@ $end_info$
 
 #pragma once
 
-#include "Interface/Core/LookupCache.h"
 #include "Interface/Core/BlockSamplingData.h"
 #include "Interface/Core/Dispatcher/Dispatcher.h"
 
@@ -260,6 +259,7 @@ private:
   DEF_OP(AtomicFetchAnd);
   DEF_OP(AtomicFetchOr);
   DEF_OP(AtomicFetchXor);
+  DEF_OP(AtomicFetchNeg);
 
   ///< Branch ops
   DEF_OP(GuestCallDirect);

External/FEXCore/Source/Interface/Core/JIT/x86_64/MemoryOps.cpp

@@ -5,9 +5,15 @@ $end_info$
 */
 
 #include "Interface/Core/JIT/x86_64/JITClass.h"
-#include "Interface/IR/Passes/RegisterAllocationPass.h"
 
-#include <cmath>
+#include <FEXCore/Core/CoreState.h>
+#include <FEXCore/IR/IR.h>
+#include <FEXCore/Utils/LogManager.h>
+
+#include <array>
+#include <stddef.h>
+#include <stdint.h>
+#include <xbyak/xbyak.h>
 
 namespace FEXCore::CPU {
 
@@ -36,10 +42,10 @@ DEF_OP(LoadContext) {
     }
     break;
     case 16: {
-      LOGMAN_MSG_A("Invalid GPR load of size 16");
+      LOGMAN_MSG_A_FMT("Invalid GPR load of size 16");
     }
     break;
-    default:  LOGMAN_MSG_A("Unhandled LoadContext size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled LoadContext size: {}", OpSize);
     }
   }
   else {
@@ -69,7 +75,7 @@ DEF_OP(LoadContext) {
         movups(GetDst(Node), xword [STATE + Op->Offset]);
     }
     break;
-    default:  LOGMAN_MSG_A("Unhandled LoadContext size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled LoadContext size: {}", OpSize);
     }
   }
 }
@@ -98,9 +104,9 @@ DEF_OP(StoreContext) {
     }
     break;
     case 16:
-      LogMan::Msg::D("Invalid store size of 16");
+      LogMan::Msg::DFmt("Invalid store size of 16");
     break;
-    default:  LOGMAN_MSG_A("Unhandled StoreContext size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled StoreContext size: {}", OpSize);
     }
   }
   else {
@@ -129,7 +135,7 @@ DEF_OP(StoreContext) {
         movups(xword [STATE + Op->Offset], GetSrc(Op->Header.Args[0].ID()));
     }
     break;
-    default:  LOGMAN_MSG_A("Unhandled StoreContext size: %d", OpSize);
+    default:  LOGMAN_MSG_A_FMT("Unhandled StoreContext size: {}", OpSize);
     }
   }
 }
@@ -160,17 +166,18 @@ DEF_OP(LoadContextIndexed) {
         mov(GetDst<RA_64>(Node),  qword [rax + index * Op->Stride]);
         break;
       default:
-        LOGMAN_MSG_A("Unhandled LoadContextIndexed size: %d", Op->Size);
+        LOGMAN_MSG_A_FMT("Unhandled LoadContextIndexed size: {}", Op->Size);
+        break;
       }
       break;
     }
     case 16:
-      LOGMAN_MSG_A("Invalid Class load of size 16");
+      LOGMAN_MSG_A_FMT("Invalid Class load of size 16");
       break;
     default:
-      LOGMAN_MSG_A("Unhandled LoadContextIndexed stride: %d", Op->Stride);
+      LOGMAN_MSG_A_FMT("Unhandled LoadContextIndexed stride: {}", Op->Stride);
+      break;
     }
-
   }
   else {
     switch (Op->Stride) {
@@ -195,7 +202,8 @@ DEF_OP(LoadContextIndexed) {
         vmovq(GetDst(Node),  qword [rax + index * Op->Stride]);
         break;
       default:
-        LOGMAN_MSG_A("Unhandled LoadContextIndexed size: %d", Op->Size);
+        LOGMAN_MSG_A_FMT("Unhandled LoadContextIndexed size: {}", Op->Size);
+        break;
       }
       break;
     }
@@ -223,12 +231,14 @@ DEF_OP(LoadContextIndexed) {
           movups(GetDst(Node), xword [STATE + rax]);
         break;
       default:
-        LOGMAN_MSG_A("Unhandled LoadContextIndexed size: %d", Op->Size);
+        LOGMAN_MSG_A_FMT("Unhandled LoadContextIndexed size: {}", Op->Size);
+        break;
       }
       break;
     }
     default:
-      LOGMAN_MSG_A("Unhandled LoadContextIndexed stride: %d", Op->Stride);
+      LOGMAN_MSG_A_FMT("Unhandled LoadContextIndexed stride: {}", Op->Stride);
+      break;
     }
   }
 }
@@ -248,13 +258,14 @@ DEF_OP(StoreContextIndexed) {
     case 4:
     case 8: {
       if (!(size == 1 || size == 2 || size == 4 || size == 8)) {
-        LOGMAN_MSG_A("Unhandled StoreContextIndexed size: %d", Op->Size);
+        LOGMAN_MSG_A_FMT("Unhandled StoreContextIndexed size: {}", Op->Size);
       }
       mov(AddressFrame(Op->Size * 8) [rax + index * Op->Stride], value);
       break;
     }
     default:
-      LOGMAN_MSG_A("Unhandled StoreContextIndexed stride: %d", Op->Stride);
+      LOGMAN_MSG_A_FMT("Unhandled StoreContextIndexed stride: {}", Op->Stride);
+      break;
     }
   }
   else {
@@ -279,7 +290,8 @@ DEF_OP(StoreContextIndexed) {
         vmovq(AddressFrame(Op->Size * 8) [rax + index * Op->Stride], value);
         break;
       default:
-        LOGMAN_MSG_A("Unhandled StoreContextIndexed size: %d", size);
+        LOGMAN_MSG_A_FMT("Unhandled StoreContextIndexed size: {}", size);
+        break;
       }
       break;
     }
@@ -307,12 +319,14 @@ DEF_OP(StoreContextIndexed) {
           movups(xword [STATE + rax], value);
         break;
       default:
-        LOGMAN_MSG_A("Unhandled StoreContextIndexed size: %d", size);
+        LOGMAN_MSG_A_FMT("Unhandled StoreContextIndexed size: {}", size);
+        break;
       }
       break;
     }
     default:
-      LOGMAN_MSG_A("Unhandled StoreContextIndexed stride: %d", Op->Stride);
+      LOGMAN_MSG_A_FMT("Unhandled StoreContextIndexed stride: {}", Op->Stride);
+      break;
     }
   }
 }
@@ -340,7 +354,7 @@ DEF_OP(SpillRegister) {
         mov(qword [rsp + SlotOffset], GetSrc<RA_64>(Op->Header.Args[0].ID()));
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled SpillRegister size: %d", OpSize);
+      default:  LOGMAN_MSG_A_FMT("Unhandled SpillRegister size: {}", OpSize);
     }
   } else if (Op->Class == FEXCore::IR::FPRClass) {
     switch (OpSize) {
@@ -356,10 +370,10 @@ DEF_OP(SpillRegister) {
         movaps(xword [rsp + SlotOffset], GetSrc(Op->Header.Args[0].ID()));
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled SpillRegister size: %d", OpSize);
+      default:  LOGMAN_MSG_A_FMT("Unhandled SpillRegister size: {}", OpSize);
     }
   } else {
-    LOGMAN_MSG_A("Unhandled SpillRegister class: %d", Op->Class.Val);
+    LOGMAN_MSG_A_FMT("Unhandled SpillRegister class: {}", Op->Class.Val);
   }
 
 
@@ -388,7 +402,7 @@ DEF_OP(FillRegister) {
         mov(GetDst<RA_64>(Node), qword [rsp + SlotOffset]);
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled FillRegister size: %d", OpSize);
+      default:  LOGMAN_MSG_A_FMT("Unhandled FillRegister size: {}", OpSize);
     }
   } else if (Op->Class == FEXCore::IR::FPRClass) {
     switch (OpSize) {
@@ -404,10 +418,10 @@ DEF_OP(FillRegister) {
         movaps(GetDst(Node), xword [rsp + SlotOffset]);
         break;
       }
-      default:  LOGMAN_MSG_A("Unhandled FillRegister size: %d", OpSize);
+      default:  LOGMAN_MSG_A_FMT("Unhandled FillRegister size: {}", OpSize);
     }
   } else {
-    LOGMAN_MSG_A("Unhandled FillRegister class: %d", Op->Class.Val);
+    LOGMAN_MSG_A_FMT("Unhandled FillRegister class: {}", Op->Class.Val);
   }
 }
 
@@ -430,11 +444,11 @@ Xbyak::RegExp X86JITCore::GenerateModRM(Xbyak::Reg Base, IR::OrderedNodeWrapper
     return Base;
   } else {
     if (OffsetScale != 1 && OffsetScale != 2 && OffsetScale != 4 && OffsetScale != 8) {
-      LOGMAN_MSG_A("Unhandled GenerateModRM OffsetScale: %d", OffsetScale);
+      LOGMAN_MSG_A_FMT("Unhandled GenerateModRM OffsetScale: {}", OffsetScale);
     }
 
     if (OffsetType != IR::MEM_OFFSET_SXTX) {
-      LOGMAN_MSG_A("Unhandled GenerateModRM OffsetType: %d", OffsetType.Val);
+      LOGMAN_MSG_A_FMT("Unhandled GenerateModRM OffsetType: {}", OffsetType.Val);
     }
 
     uint64_t Const;
@@ -475,7 +489,7 @@ DEF_OP(LoadMem) {
         mov(Dst, qword [MemPtr]);
       }
       break;
-      default:  LOGMAN_MSG_A("Unhandled LoadMem size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled LoadMem size: {}", Op->Size);
     }
   }
   else
@@ -511,7 +525,7 @@ DEF_OP(LoadMem) {
          }
        }
        break;
-      default:  LOGMAN_MSG_A("Unhandled LoadMem size: %d", Op->Size);
+      default:  LOGMAN_MSG_A_FMT("Unhandled LoadMem size: {}", Op->Size);
     }
   }
 }
@@ -537,7 +551,7 @@ DEF_OP(StoreMem) {
     case 8:
       mov(qword [MemPtr], GetSrc<RA_64>(Op->Header.Args[1].ID()));
     break;
-    default:  LOGMAN_MSG_A("Unhandled StoreMem size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled StoreMem size: {}", Op->Size);
     }
   }
   else {
@@ -560,17 +574,17 @@ DEF_OP(StoreMem) {
       else
         movups(xword [MemPtr], GetSrc(Op->Header.Args[1].ID()));
     break;
-    default:  LOGMAN_MSG_A("Unhandled StoreMem size: %d", Op->Size);
+    default:  LOGMAN_MSG_A_FMT("Unhandled StoreMem size: {}", Op->Size);
     }
   }
 }
 
 DEF_OP(VLoadMemElement) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(VStoreMemElement) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(CacheLineClear) {

External/FEXCore/Source/Interface/Core/JIT/x86_64/MiscOps.cpp

@@ -4,16 +4,26 @@ tags: backend|x86-64
 $end_info$
 */
 
+#include "Interface/Context/Context.h"
+#include "Interface/Core/Dispatcher/Dispatcher.h"
 #include "Interface/Core/JIT/x86_64/JITClass.h"
-#include "Interface/IR/Passes/RegisterAllocationPass.h"
+
+#include <FEXCore/Core/CoreState.h>
+#include <FEXCore/Utils/LogManager.h>
+#include <FEXCore/IR/IR.h>
+
+#include <array>
+#include <stddef.h>
+#include <stdint.h>
+#include <xbyak/xbyak.h>
 
 namespace FEXCore::CPU {
 static void PrintValue(uint64_t Value) {
-  LogMan::Msg::D("Value: 0x%lx", Value);
+  LogMan::Msg::DFmt("Value: 0x{:x}", Value);
 }
 
 static void PrintVectorValue(uint64_t Value, uint64_t ValueUpper) {
-  LogMan::Msg::D("Value: 0x%016lx'%016lx", ValueUpper, Value);
+  LogMan::Msg::DFmt("Value: 0x{:016x}'{:016x}", ValueUpper, Value);
 }
 
 #define DEF_OP(x) void X86JITCore::Op_##x(FEXCore::IR::IROp_Header *IROp, uint32_t Node)
@@ -30,7 +40,7 @@ DEF_OP(Fence) {
     case IR::Fence_Store.Val:
       sfence();
       break;
-    default: LOGMAN_MSG_A("Unknown Fence: %d", Op->Fence); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Fence: {}", Op->Fence); break;
   }
 }
 
@@ -83,7 +93,7 @@ DEF_OP(Break) {
       }
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Break reason: %d", Op->Reason);
+    default: LOGMAN_MSG_A_FMT("Unknown Break reason: {}", Op->Reason);
   }
 }
 

External/FEXCore/Source/Interface/Core/JIT/x86_64/MoveOps.cpp

@@ -5,7 +5,13 @@ $end_info$
 */
 
 #include "Interface/Core/JIT/x86_64/JITClass.h"
-#include "Interface/IR/Passes/RegisterAllocationPass.h"
+
+#include <FEXCore/Utils/LogManager.h>
+#include <FEXCore/IR/IR.h>
+
+#include <array>
+#include <stdint.h>
+#include <utility>
 
 namespace FEXCore::CPU {
 
@@ -25,7 +31,7 @@ DEF_OP(ExtractElementPair) {
       mov (GetDst<RA_64>(Node), Regs[Op->Element]);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Size"); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Size"); break;
   }
 }
 
@@ -51,7 +57,7 @@ DEF_OP(CreateElementPair) {
       RegTmp = rax;
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Size"); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Size"); break;
   }
 
   if (Dst.first != RegSecond) {

External/FEXCore/Source/Interface/Core/JIT/x86_64/VectorOps.cpp

@@ -5,8 +5,14 @@ $end_info$
 */
 
 #include "Interface/Core/JIT/x86_64/JITClass.h"
-#include "Interface/IR/Passes/RegisterAllocationPass.h"
 
+#include <FEXCore/IR/IR.h>
+#include <FEXCore/Utils/LogManager.h>
+
+#include <array>
+#include <stddef.h>
+#include <stdint.h>
+#include <xbyak/xbyak.h>
 
 namespace FEXCore::CPU {
 
@@ -62,24 +68,24 @@ DEF_OP(VectorImm) {
 }
 
 DEF_OP(CreateVector2) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(CreateVector4) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(SplatVector) {
   auto Op = IROp->C<IR::IROp_SplatVector2>();
   uint8_t OpSize = IROp->Size;
 
-  LOGMAN_THROW_A(OpSize <= 16, "Can't handle a vector of size: %d", OpSize);
+  LOGMAN_THROW_A_FMT(OpSize <= 16, "Can't handle a vector of size: {}", OpSize);
   uint8_t Elements = 0;
 
   switch (Op->Header.Op) {
     case IR::OP_SPLATVECTOR4: Elements = 4; break;
     case IR::OP_SPLATVECTOR2: Elements = 2; break;
-    default: LOGMAN_MSG_A("Uknown Splat size"); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Splat size"); break;
   }
 
   uint8_t ElementSize = OpSize / Elements;
@@ -92,7 +98,7 @@ DEF_OP(SplatVector) {
     case 8:
       movddup(GetDst(Node), GetSrc(Op->Header.Args[0].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.Size); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.Size); break;
   }
 }
 
@@ -130,7 +136,7 @@ DEF_OP(VMov) {
       movaps(GetDst(Node), GetSrc(Op->Header.Args[0].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", OpSize); break;
   }
 }
 
@@ -176,7 +182,7 @@ DEF_OP(VAdd) {
       vpaddq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -199,7 +205,7 @@ DEF_OP(VSub) {
       vpsubq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -214,7 +220,7 @@ DEF_OP(VUQAdd) {
       vpaddusw(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -229,7 +235,7 @@ DEF_OP(VUQSub) {
       vpsubusw(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -244,7 +250,7 @@ DEF_OP(VSQAdd) {
       vpaddsw(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -259,7 +265,7 @@ DEF_OP(VSQSub) {
       vpsubsw(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -296,7 +302,7 @@ DEF_OP(VAddP) {
       case 4:
         vphaddd(GetDst(Node), xmm15, xmm14);
         break;
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -325,7 +331,7 @@ DEF_OP(VAddP) {
       case 4:
         vphaddd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
         break;
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -357,7 +363,7 @@ DEF_OP(VAddV) {
       pinsrd(xmm15, eax, 0);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 
   movaps(Dest, xmm15);
@@ -376,7 +382,7 @@ DEF_OP(VUMinV) {
       pinsrw(Dest, eax, 1);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -391,7 +397,7 @@ DEF_OP(VURAvg) {
       vpavgw(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -414,7 +420,7 @@ DEF_OP(VAbs) {
       vpabsq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -438,7 +444,7 @@ DEF_OP(VPopcount) {
       }
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 
   movaps(Dest, xmm15);
@@ -459,7 +465,7 @@ DEF_OP(VFAdd) {
         vaddsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -473,7 +479,7 @@ DEF_OP(VFAdd) {
         vaddpd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -487,7 +493,7 @@ DEF_OP(VFAddP) {
     case 8:
       vhaddpd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -506,7 +512,7 @@ DEF_OP(VFSub) {
         vsubsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -520,7 +526,7 @@ DEF_OP(VFSub) {
         vsubpd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -540,7 +546,7 @@ DEF_OP(VFMul) {
         vmulsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -554,7 +560,7 @@ DEF_OP(VFMul) {
         vmulpd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -574,7 +580,7 @@ DEF_OP(VFDiv) {
         vdivsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -588,7 +594,7 @@ DEF_OP(VFDiv) {
         vdivpd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -608,7 +614,7 @@ DEF_OP(VFMin) {
         vminsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -622,7 +628,7 @@ DEF_OP(VFMin) {
         vminpd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -642,7 +648,7 @@ DEF_OP(VFMax) {
         vmaxsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -656,7 +662,7 @@ DEF_OP(VFMax) {
         vmaxpd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -674,7 +680,7 @@ DEF_OP(VFRecp) {
         vdivss(GetDst(Node), xmm15, GetSrc(Op->Header.Args[0].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -687,7 +693,7 @@ DEF_OP(VFRecp) {
         vdivps(GetDst(Node), xmm15, GetSrc(Op->Header.Args[0].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -707,7 +713,7 @@ DEF_OP(VFSqrt) {
         vsqrtsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[0].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -721,7 +727,7 @@ DEF_OP(VFSqrt) {
         vsqrtpd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()));
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -747,7 +753,7 @@ DEF_OP(VFRSqrt) {
         divsd(GetDst(Node), xmm15);
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -761,7 +767,7 @@ DEF_OP(VFRSqrt) {
         divps(GetDst(Node), xmm15);
       break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -786,7 +792,7 @@ DEF_OP(VNeg) {
       vpsubq(GetDst(Node), xmm15, GetSrc(Op->Header.Args[0].ID()));
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -807,7 +813,7 @@ DEF_OP(VFNeg) {
       vxorpd(GetDst(Node), xmm15, GetSrc(Op->Header.Args[0].ID()));
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -830,7 +836,7 @@ DEF_OP(VUMin) {
         pinsrq(GetDst(Node), TMP2, 0);
         break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -847,7 +853,7 @@ DEF_OP(VUMin) {
         vpminud(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
         break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -867,7 +873,7 @@ DEF_OP(VSMin) {
       vpminsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -886,7 +892,7 @@ DEF_OP(VUMax) {
       vpmaxud(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -905,7 +911,7 @@ DEF_OP(VSMax) {
       vpmaxsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -930,7 +936,7 @@ DEF_OP(VZip) {
       punpcklqdq(xmm15, GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
   movapd(GetDst(Node), xmm15);
 }
@@ -957,7 +963,7 @@ DEF_OP(VZip2) {
       vpunpckhdq(GetDst(Node), xmm15, xmm14);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
   else {
@@ -978,7 +984,7 @@ DEF_OP(VZip2) {
       punpckhqdq(xmm15, GetSrc(Op->Header.Args[1].ID()));
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
     movapd(GetDst(Node), xmm15);
   }
@@ -989,7 +995,7 @@ DEF_OP(VUnZip) {
   uint8_t OpSize = IROp->Size;
 
   if (OpSize == 8) {
-    LOGMAN_MSG_A("Unsupported registersize on VunZip");
+    LOGMAN_MSG_A_FMT("Unsupported register size on VUnZip");
   }
   else {
     switch (Op->Header.ElementSize) {
@@ -1031,7 +1037,7 @@ DEF_OP(VUnZip) {
           0b0'0);
         break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1042,7 +1048,7 @@ DEF_OP(VUnZip2) {
 
 
   if (OpSize == 8) {
-    LOGMAN_MSG_A("Unsupported registersize on VunZip");
+    LOGMAN_MSG_A_FMT("Unsupported register size on VUnZip2");
   }
   else {
     switch (Op->Header.ElementSize) {
@@ -1084,7 +1090,7 @@ DEF_OP(VUnZip2) {
           0b1'1);
         break;
       }
-      default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+      default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
     }
   }
 }
@@ -1113,7 +1119,7 @@ DEF_OP(VCMPEQ) {
     case 8:
       vpcmpeqq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -1134,7 +1140,7 @@ DEF_OP(VCMPEQZ) {
     case 8:
       vpcmpeqq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), xmm15);
       break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -1154,7 +1160,7 @@ DEF_OP(VCMPGT) {
     case 8:
       vpcmpgtq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -1175,7 +1181,7 @@ DEF_OP(VCMPGTZ) {
     case 8:
       vpcmpgtq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), xmm15);
       break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -1196,7 +1202,7 @@ DEF_OP(VCMPLTZ) {
     case 8:
       vpcmpgtq(GetDst(Node), xmm15, GetSrc(Op->Header.Args[0].ID()));
       break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -1212,7 +1218,7 @@ DEF_OP(VFCMPEQ) {
     case 8:
       vcmpsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()), 0);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
   else {
@@ -1223,7 +1229,7 @@ DEF_OP(VFCMPEQ) {
     case 8:
       vcmppd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()), 0);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
 }
@@ -1240,7 +1246,7 @@ DEF_OP(VFCMPNEQ) {
     case 8:
       vcmpsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()), 4);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
 
   }
@@ -1252,7 +1258,7 @@ DEF_OP(VFCMPNEQ) {
     case 8:
       vcmppd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()), 4);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
 }
@@ -1269,7 +1275,7 @@ DEF_OP(VFCMPLT) {
     case 8:
       vcmpsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()), 1);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
   else {
@@ -1280,7 +1286,7 @@ DEF_OP(VFCMPLT) {
     case 8:
       vcmppd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()), 1);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
 }
@@ -1297,7 +1303,7 @@ DEF_OP(VFCMPGT) {
     case 8:
       vcmpsd(GetDst(Node), GetSrc(Op->Header.Args[1].ID()), GetSrc(Op->Header.Args[0].ID()), 1);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
   else {
@@ -1308,7 +1314,7 @@ DEF_OP(VFCMPGT) {
     case 8:
       vcmppd(GetDst(Node), GetSrc(Op->Header.Args[1].ID()), GetSrc(Op->Header.Args[0].ID()), 1);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
 }
@@ -1325,7 +1331,7 @@ DEF_OP(VFCMPLE) {
     case 8:
       vcmpsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()), 2);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
   else {
@@ -1336,7 +1342,7 @@ DEF_OP(VFCMPLE) {
     case 8:
       vcmppd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()), 2);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
 }
@@ -1353,7 +1359,7 @@ DEF_OP(VFCMPORD) {
     case 8:
       vcmpsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()), 7);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
   else {
@@ -1364,7 +1370,7 @@ DEF_OP(VFCMPORD) {
     case 8:
       vcmppd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()), 7);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
 }
@@ -1381,7 +1387,7 @@ DEF_OP(VFCMPUNO) {
     case 8:
       vcmpsd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()), 3);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
   else {
@@ -1392,21 +1398,21 @@ DEF_OP(VFCMPUNO) {
     case 8:
       vcmppd(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()), 3);
     break;
-    default: LOGMAN_MSG_A("Unsupported elementSize: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unsupported element size: {}", Op->Header.ElementSize);
     }
   }
 }
 
 DEF_OP(VUShl) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(VUShr) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(VSShr) {
-  LOGMAN_MSG_A("Unimplemented");
+  LOGMAN_MSG_A_FMT("Unimplemented");
 }
 
 DEF_OP(VUShlS) {
@@ -1425,7 +1431,7 @@ DEF_OP(VUShlS) {
       vpsllq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1445,7 +1451,7 @@ DEF_OP(VUShrS) {
       vpsrlq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1462,7 +1468,7 @@ DEF_OP(VSShrS) {
       break;
     }
     case 8: // Doesn't exist on x86
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1495,7 +1501,7 @@ DEF_OP(VInsElement) {
     pinsrq(xmm15, rax, Op->DestIdx);
   break;
   }
-  default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+  default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 
   movapd(GetDst(Node), xmm15);
@@ -1530,7 +1536,7 @@ DEF_OP(VInsScalarElement) {
     pinsrq(xmm15, rax, Op->DestIdx);
   break;
   }
-  default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+  default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 
   movapd(GetDst(Node), xmm15);
@@ -1560,7 +1566,7 @@ DEF_OP(VExtractElement) {
       pinsrq(GetDst(Node), rax, 0);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1613,7 +1619,7 @@ DEF_OP(VDupElement) {
         (Op->Index << 1));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1666,7 +1672,7 @@ DEF_OP(VUShrI) {
       psrlq(GetDst(Node), Op->BitShift);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1704,7 +1710,7 @@ DEF_OP(VSShrI) {
       }
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1724,7 +1730,7 @@ DEF_OP(VShlI) {
       psllq(GetDst(Node), Op->BitShift);
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1754,7 +1760,7 @@ DEF_OP(VUShrNI) {
       mov(rcx, 0x80'80'80'80'80'80'80'80); // Upper
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 
   vmovq(xmm15, rax);
@@ -1790,7 +1796,7 @@ DEF_OP(VUShrNI2) {
       mov(rcx, 0x0B'0A'09'08'03'02'01'00); // Upper
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 
   vmovq(xmm15, rax);
@@ -1817,7 +1823,7 @@ DEF_OP(VSXTL) {
     case 8:
       pmovsxdq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -1836,7 +1842,7 @@ DEF_OP(VSXTL2) {
     case 8:
       pmovsxdq(GetDst(Node), GetDst(Node));
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -1852,7 +1858,7 @@ DEF_OP(VUXTL) {
     case 8:
       pmovzxdq(GetDst(Node), GetSrc(Op->Header.Args[0].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -1871,7 +1877,7 @@ DEF_OP(VUXTL2) {
     case 8:
       pmovzxdq(GetDst(Node), GetDst(Node));
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
 }
 
@@ -1884,7 +1890,7 @@ DEF_OP(VSQXTN) {
     case 2:
       packssdw(xmm15, GetSrc(Op->Header.Args[0].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
   psrldq(xmm15, 8);
   movaps(GetDst(Node), xmm15);
@@ -1903,7 +1909,7 @@ DEF_OP(VSQXTN2) {
     case 2:
       packssdw(xmm15, GetSrc(Op->Header.Args[1].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
 
   if (OpSize == 8) {
@@ -1921,7 +1927,7 @@ DEF_OP(VSQXTUN) {
     case 2:
       packusdw(xmm15, GetSrc(Op->Header.Args[0].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
   psrldq(xmm15, 8);
   movaps(GetDst(Node), xmm15);
@@ -1940,7 +1946,7 @@ DEF_OP(VSQXTUN2) {
     case 2:
       packusdw(xmm15, GetSrc(Op->Header.Args[1].ID()));
     break;
-    default: LOGMAN_MSG_A("Unknown element size: %d", Op->Header.ElementSize);
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize);
   }
   if (OpSize == 8) {
     psrldq(xmm15, OpSize / 2);
@@ -1960,7 +1966,7 @@ DEF_OP(VMul) {
       vpmulld(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -1991,7 +1997,7 @@ DEF_OP(VUMull) {
       vpmuludq(GetDst(Node), xmm14, xmm15);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -2026,7 +2032,7 @@ DEF_OP(VSMull) {
       vpmuldq(GetDst(Node), xmm14, xmm15);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -2062,7 +2068,7 @@ DEF_OP(VUMull2) {
       vpmuludq(GetDst(Node), xmm14, xmm15);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -2102,7 +2108,7 @@ DEF_OP(VSMull2) {
       vpmuldq(GetDst(Node), xmm14, xmm15);
     break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -2123,7 +2129,7 @@ DEF_OP(VUABDL) {
       vpabsd(GetDst(Node), GetDst(Node));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown Element Size: %d", Op->Header.ElementSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown Element Size: {}", Op->Header.ElementSize); break;
   }
 }
 
@@ -2141,7 +2147,7 @@ DEF_OP(VTBL1) {
       vpshufb(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[1].ID()));
       break;
     }
-    default: LOGMAN_MSG_A("Unknown OpSize: %d", OpSize); break;
+    default: LOGMAN_MSG_A_FMT("Unknown OpSize: {}", OpSize); break;
   }
 }
 

External/FEXCore/Source/Interface/Core/LookupCache.cpp

@@ -5,10 +5,11 @@ desc: Stores information about blocks, and provides C++ implementations to looku
 $end_info$
 */
 
-#include "Interface/Context/Context.h"
-#include "Interface/Core/Core.h"
-#include "Interface/Core/LookupCache.h"
 #include <FEXCore/Utils/Allocator.h>
+#include <FEXCore/Utils/LogManager.h>
+
+#include "Interface/Context/Context.h"
+#include "Interface/Core/LookupCache.h"
 
 #include <sys/mman.h>
 

External/FEXCore/Source/Interface/Core/LookupCache.h

@@ -1,10 +1,18 @@
 #pragma once
-#include "Interface/Context/Context.h"
 #include <FEXCore/Utils/LogManager.h>
 
+#include <cstdint>
+#include <functional>
 #include <map>
+#include <stddef.h>
+#include <utility>
+#include <vector>
 
 namespace FEXCore {
+namespace Context {
+  struct Context;
+}
+
 class LookupCache {
 public:
 

External/FEXCore/Source/Interface/Core/OpcodeDispatcher.h

@@ -3,7 +3,8 @@
 #include "Interface/Core/Frontend.h"
 #include "Interface/Context/Context.h"
 
-#include <FEXCore/Core/CoreState.h>
+#include <FEXCore/Config/Config.h>
+#include <FEXCore/Core/Context.h>
 #include <FEXCore/Debug/X86Tables.h>
 #include <FEXCore/IR/IntrusiveIRList.h>
 #include <FEXCore/IR/IR.h>
@@ -12,9 +13,10 @@
 #include <FEXCore/Utils/LogManager.h>
 
 #include <cstdint>
-#include <functional>
 #include <map>
-#include <set>
+#include <stddef.h>
+#include <utility>
+#include <vector>
 
 namespace FEXCore::IR {
 class Pass;
@@ -270,15 +272,15 @@ public:
   template<size_t ElementSize, bool HalfSize, bool Low>
   void PSHUFDOp(OpcodeArgs);
   void MOVDOp(OpcodeArgs);
-  template<size_t ElementSize, bool Scalar, uint32_t SrcIndex>
+  template<size_t ElementSize>
   void PSRLDOp(OpcodeArgs);
   template<size_t ElementSize>
   void PSRLI(OpcodeArgs);
   template<size_t ElementSize>
   void PSLLI(OpcodeArgs);
-  template<size_t ElementSize, bool Scalar, uint32_t SrcIndex>
+  template<size_t ElementSize>
   void PSLL(OpcodeArgs);
-  template<size_t ElementSize, bool Scalar, uint32_t SrcIndex>
+  template<size_t ElementSize>
   void PSRAOp(OpcodeArgs);
   void PSRLDQ(OpcodeArgs);
   void PSLLDQ(OpcodeArgs);
@@ -299,9 +301,9 @@ public:
   void Vector_CVT_Float_To_Float(OpcodeArgs);
   template<size_t SrcElementSize, bool Narrow, bool HostRoundingMode>
   void Vector_CVT_Float_To_Int(OpcodeArgs);
-  template<size_t SrcElementSize, bool Signed, bool Widen>
+  template<size_t SrcElementSize, bool Widen>
   void MMX_To_XMM_Vector_CVT_Int_To_Float(OpcodeArgs);
-  template<size_t SrcElementSize, bool Narrow, bool HostRoundingMode>
+  template<size_t SrcElementSize, bool HostRoundingMode>
   void XMM_To_MMX_Vector_CVT_Float_To_Int(OpcodeArgs);
   void MASKMOVOp(OpcodeArgs);
   void MOVBetweenGPR_FPR(OpcodeArgs);
@@ -380,6 +382,8 @@ public:
   void X87FRSTOR(OpcodeArgs);
   void X87FXAM(OpcodeArgs);
   void X87FCMOV(OpcodeArgs);
+  void X87EMMS(OpcodeArgs);
+  void X87FFREE(OpcodeArgs);
 
   void FXCH(OpcodeArgs);
 
@@ -501,9 +505,19 @@ private:
   uint8_t GetSrcSize(FEXCore::X86Tables::DecodedOp Op) const;
 
   template<unsigned BitOffset>
-  void SetRFLAG(OrderedNode *Value);
-  void SetRFLAG(OrderedNode *Value, unsigned BitOffset);
-  OrderedNode *GetRFLAG(unsigned BitOffset);
+  void SetRFLAG(OrderedNode *Value) {
+    flagsOp = FLAGS_OP_NONE;
+    _StoreFlag(_Bfe(1, 0, Value), BitOffset);
+  }
+
+  void SetRFLAG(OrderedNode *Value, unsigned BitOffset) {
+    flagsOp = FLAGS_OP_NONE;
+    _StoreFlag(_Bfe(1, 0, Value), BitOffset);
+  }
+
+  OrderedNode *GetRFLAG(unsigned BitOffset) {
+    return _LoadFlag(BitOffset);
+  }
 
   OrderedNode *SelectCC(uint8_t OP, OrderedNode *TrueValue, OrderedNode *FalseValue);
 
@@ -526,6 +540,14 @@ private:
   void GenerateFlags_RotateLeftImmediate(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, uint64_t Shift);
 
   OrderedNode * GetX87Top();
+  enum X87Tag {
+    TAG_VALID   = 0b00,
+    TAG_ZERO    = 0b01,
+    TAG_SPECIAL = 0b10,
+    TAG_EMPTY   = 0b11
+  };
+  void SetX87TopTag(OrderedNode *Value, uint32_t Tag);
+  OrderedNode *GetX87FTW(OrderedNode *Value);
   void SetX87Top(OrderedNode *Value);
 
   bool DestIsLockedMem(FEXCore::X86Tables::DecodedOp Op) const {

External/FEXCore/Source/Interface/Core/OpcodeDispatcher/Crypto.cpp

@@ -0,0 +1,63 @@
+/*
+$info$
+tags: frontend|x86-to-ir, opcodes|dispatcher-implementations
+desc: Handles x86/64 Crypto instructions to IR
+$end_info$
+*/
+
+#include <FEXCore/Debug/X86Tables.h>
+#include <FEXCore/IR/IREmitter.h>
+#include <FEXCore/Utils/LogManager.h>
+#include "Interface/Core/OpcodeDispatcher.h"
+
+#include <stdint.h>
+
+namespace FEXCore::IR {
+class OrderedNode;
+
+#define OpcodeArgs [[maybe_unused]] FEXCore::X86Tables::DecodedOp Op
+
+void OpDispatchBuilder::AESImcOp(OpcodeArgs) {
+  OrderedNode *Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, -1);
+  auto Res = _VAESImc(Src);
+  StoreResult(FPRClass, Op, Res, -1);
+}
+
+void OpDispatchBuilder::AESEncOp(OpcodeArgs) {
+  OrderedNode *Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags, -1);
+  OrderedNode *Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, -1);
+  auto Res = _VAESEnc(Dest, Src);
+  StoreResult(FPRClass, Op, Res, -1);
+}
+
+void OpDispatchBuilder::AESEncLastOp(OpcodeArgs) {
+  OrderedNode *Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags, -1);
+  OrderedNode *Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, -1);
+  auto Res = _VAESEncLast(Dest, Src);
+  StoreResult(FPRClass, Op, Res, -1);
+}
+
+void OpDispatchBuilder::AESDecOp(OpcodeArgs) {
+  OrderedNode *Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags, -1);
+  OrderedNode *Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, -1);
+  auto Res = _VAESDec(Dest, Src);
+  StoreResult(FPRClass, Op, Res, -1);
+}
+
+void OpDispatchBuilder::AESDecLastOp(OpcodeArgs) {
+  OrderedNode *Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags, -1);
+  OrderedNode *Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, -1);
+  auto Res = _VAESDecLast(Dest, Src);
+  StoreResult(FPRClass, Op, Res, -1);
+}
+
+void OpDispatchBuilder::AESKeyGenAssist(OpcodeArgs) {
+  OrderedNode *Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, -1);
+  LOGMAN_THROW_A(Op->Src[1].IsLiteral(), "Src1 needs to be literal here");
+  uint64_t RCON = Op->Src[1].Data.Literal.Value;
+
+  auto Res = _VAESKeyGenAssist(Src, RCON);
+  StoreResult(FPRClass, Op, Res, -1);
+}
+
+}

External/FEXCore/Source/Interface/Core/OpcodeDispatcher/Flags.cpp

@@ -0,0 +1,818 @@
+/*
+$info$
+tags: frontend|x86-to-ir, opcodes|dispatcher-implementations
+desc: Handles x86/64 flag generation
+$end_info$
+*/
+
+#include "Interface/Context/Context.h"
+#include "Interface/Core/OpcodeDispatcher.h"
+
+#include <FEXCore/Core/X86Enums.h>
+#include <FEXCore/Config/Config.h>
+#include <FEXCore/Debug/X86Tables.h>
+#include <FEXCore/Utils/LogManager.h>
+#include <FEXCore/IR/IR.h>
+
+#include <array>
+#include <cstdint>
+
+namespace FEXCore::IR {
+constexpr std::array<uint32_t, 17> FlagOffsets = {
+  FEXCore::X86State::RFLAG_CF_LOC,
+  FEXCore::X86State::RFLAG_PF_LOC,
+  FEXCore::X86State::RFLAG_AF_LOC,
+  FEXCore::X86State::RFLAG_ZF_LOC,
+  FEXCore::X86State::RFLAG_SF_LOC,
+  FEXCore::X86State::RFLAG_TF_LOC,
+  FEXCore::X86State::RFLAG_IF_LOC,
+  FEXCore::X86State::RFLAG_DF_LOC,
+  FEXCore::X86State::RFLAG_OF_LOC,
+  FEXCore::X86State::RFLAG_IOPL_LOC,
+  FEXCore::X86State::RFLAG_NT_LOC,
+  FEXCore::X86State::RFLAG_RF_LOC,
+  FEXCore::X86State::RFLAG_VM_LOC,
+  FEXCore::X86State::RFLAG_AC_LOC,
+  FEXCore::X86State::RFLAG_VIF_LOC,
+  FEXCore::X86State::RFLAG_VIP_LOC,
+  FEXCore::X86State::RFLAG_ID_LOC,
+};
+
+void OpDispatchBuilder::SetPackedRFLAG(bool Lower8, OrderedNode *Src) {
+  uint8_t NumFlags = FlagOffsets.size();
+  if (Lower8) {
+    NumFlags = 5;
+  }
+  auto OneConst = _Constant(1);
+  for (int i = 0; i < NumFlags; ++i) {
+    auto Tmp = _And(_Lshr(Src, _Constant(FlagOffsets[i])), OneConst);
+    SetRFLAG(Tmp, FlagOffsets[i]);
+  }
+}
+
+OrderedNode *OpDispatchBuilder::GetPackedRFLAG(bool Lower8) {
+  OrderedNode *Original = _Constant(2);
+  uint8_t NumFlags = FlagOffsets.size();
+  if (Lower8) {
+    NumFlags = 5;
+  }
+
+  for (int i = 0; i < NumFlags; ++i) {
+    OrderedNode *Flag = _LoadFlag(FlagOffsets[i]);
+    Flag = _Bfe(4, 32, 0, Flag);
+    Flag = _Lshl(Flag, _Constant(FlagOffsets[i]));
+    Original = _Or(Original, Flag);
+  }
+  return Original;
+}
+
+void OpDispatchBuilder::GenerateFlags_ADC(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2, OrderedNode *CF) {
+  auto Size = GetSrcSize(Op) * 8;
+  // AF
+  {
+    OrderedNode *AFRes = _Xor(_Xor(Src1, Src2), Res);
+    AFRes = _Bfe(1, 4, AFRes);
+    SetRFLAG<FEXCore::X86State::RFLAG_AF_LOC>(AFRes);
+  }
+
+  // SF
+  {
+    auto SignBitConst = _Constant(GetSrcSize(Op) * 8 - 1);
+
+    auto LshrOp = _Lshr(Res, SignBitConst);
+    SetRFLAG<FEXCore::X86State::RFLAG_SF_LOC>(LshrOp);
+  }
+
+  // PF
+  if (!CTX->Config.ABINoPF) {
+    auto PopCountOp = _Popcount(_And(Res, _Constant(0xFF)));
+
+    auto XorOp = _Xor(PopCountOp, _Constant(1));
+    SetRFLAG<FEXCore::X86State::RFLAG_PF_LOC>(XorOp);
+  } else {
+    _InvalidateFlags(1UL << FEXCore::X86State::RFLAG_PF_LOC);
+  }
+
+  // ZF
+  {
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ,
+        Res, _Constant(0), _Constant(1), _Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_ZF_LOC>(SelectOp);
+  }
+
+  // CF
+  // Unsigned
+  {
+    auto SelectOpLT = _Select(FEXCore::IR::COND_ULT, Res, Src2, _Constant(1), _Constant(0));
+    auto SelectOpLE = _Select(FEXCore::IR::COND_ULE, Res, Src2, _Constant(1), _Constant(0));
+    auto SelectCF   = _Select(FEXCore::IR::COND_EQ, CF, _Constant(1), SelectOpLE, SelectOpLT);
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(SelectCF);
+  }
+
+  // OF
+  // Signed
+  {
+    auto NegOne = _Constant(~0ULL);
+    auto XorOp1 = _Xor(_Xor(Src1, Src2), NegOne);
+    auto XorOp2 = _Xor(Res, Src1);
+    OrderedNode *AndOp1 = _And(XorOp1, XorOp2);
+
+    switch (Size) {
+    case 8:
+      AndOp1 = _Bfe(1, 7, AndOp1);
+    break;
+    case 16:
+      AndOp1 = _Bfe(1, 15, AndOp1);
+    break;
+    case 32:
+      AndOp1 = _Bfe(1, 31, AndOp1);
+    break;
+    case 64:
+      AndOp1 = _Bfe(1, 63, AndOp1);
+    break;
+    default: LOGMAN_MSG_A("Unknown BFESize: %d", Size); break;
+    }
+    SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(AndOp1);
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_SBB(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2, OrderedNode *CF) {
+  // AF
+  {
+    OrderedNode *AFRes = _Xor(_Xor(Src1, Src2), Res);
+    AFRes = _Bfe(1, 4, AFRes);
+    SetRFLAG<FEXCore::X86State::RFLAG_AF_LOC>(AFRes);
+  }
+
+  // SF
+  {
+    auto SignBitConst = _Constant(GetSrcSize(Op) * 8 - 1);
+
+    auto LshrOp = _Lshr(Res, SignBitConst);
+    SetRFLAG<FEXCore::X86State::RFLAG_SF_LOC>(LshrOp);
+  }
+
+  // PF
+  if (!CTX->Config.ABINoPF) {
+    auto PopCountOp = _Popcount(_And(Res, _Constant(0xFF)));
+
+    auto XorOp = _Xor(PopCountOp, _Constant(1));
+    SetRFLAG<FEXCore::X86State::RFLAG_PF_LOC>(XorOp);
+  } else {
+    _InvalidateFlags(1UL << FEXCore::X86State::RFLAG_PF_LOC);
+  }
+
+  // ZF
+  {
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ,
+        Res, _Constant(0), _Constant(1), _Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_ZF_LOC>(SelectOp);
+  }
+
+  // CF
+  // Unsigned
+  {
+    auto SelectOpLT = _Select(FEXCore::IR::COND_UGT, Res, Src1, _Constant(1), _Constant(0));
+    auto SelectOpLE = _Select(FEXCore::IR::COND_UGE, Res, Src1, _Constant(1), _Constant(0));
+    auto SelectCF   = _Select(FEXCore::IR::COND_EQ, CF, _Constant(1), SelectOpLE, SelectOpLT);
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(SelectCF);
+  }
+
+  // OF
+  // Signed
+  {
+    auto XorOp1 = _Xor(Src1, Src2);
+    auto XorOp2 = _Xor(Res, Src1);
+    OrderedNode *AndOp1 = _And(XorOp1, XorOp2);
+
+    switch (GetSrcSize(Op)) {
+    case 1:
+      AndOp1 = _Bfe(1, 7, AndOp1);
+    break;
+    case 2:
+      AndOp1 = _Bfe(1, 15, AndOp1);
+    break;
+    case 4:
+      AndOp1 = _Bfe(1, 31, AndOp1);
+    break;
+    case 8:
+      AndOp1 = _Bfe(1, 63, AndOp1);
+    break;
+    default: LOGMAN_MSG_A("Unknown BFESize: %d", GetSrcSize(Op)); break;
+    }
+    SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(AndOp1);
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_SUB(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2, bool UpdateCF) {
+  // AF
+  {
+    OrderedNode *AFRes = _Xor(_Xor(Src1, Src2), Res);
+    AFRes = _Bfe(1, 4, AFRes);
+    SetRFLAG<FEXCore::X86State::RFLAG_AF_LOC>(AFRes);
+  }
+
+  // SF
+  {
+    auto SignBitConst = _Constant(GetSrcSize(Op) * 8 - 1);
+
+    auto LshrOp = _Lshr(Res, SignBitConst);
+    SetRFLAG<FEXCore::X86State::RFLAG_SF_LOC>(LshrOp);
+  }
+
+  // PF
+  if (!CTX->Config.ABINoPF) {
+    auto EightBitMask = _Constant(0xFF);
+    auto PopCountOp = _Popcount(_And(Res, EightBitMask));
+    auto XorOp = _Xor(PopCountOp, _Constant(1));
+    SetRFLAG<FEXCore::X86State::RFLAG_PF_LOC>(XorOp);
+  } else {
+    _InvalidateFlags(1UL << FEXCore::X86State::RFLAG_PF_LOC);
+  }
+
+  // ZF
+  {
+    auto ZeroConst = _Constant(0);
+    auto OneConst = _Constant(1);
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ,
+        Res, ZeroConst, OneConst, ZeroConst);
+    SetRFLAG<FEXCore::X86State::RFLAG_ZF_LOC>(SelectOp);
+  }
+
+  // CF
+  if (UpdateCF) {
+    auto ZeroConst = _Constant(0);
+    auto OneConst = _Constant(1);
+
+    auto SelectOp = _Select(FEXCore::IR::COND_ULT,
+        Src1, Src2, OneConst, ZeroConst);
+
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(SelectOp);
+  }
+  // OF
+  {
+    auto XorOp1 = _Xor(Src1, Src2);
+    auto XorOp2 = _Xor(Res, Src1);
+    OrderedNode *FinalAnd = _And(XorOp1, XorOp2);
+
+    FinalAnd = _Bfe(1, GetSrcSize(Op) * 8 - 1, FinalAnd);
+
+    SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(FinalAnd);
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_ADD(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2, bool UpdateCF) {
+  // AF
+  {
+    OrderedNode *AFRes = _Xor(_Xor(Src1, Src2), Res);
+    AFRes = _Bfe(1, 4, AFRes);
+    SetRFLAG<FEXCore::X86State::RFLAG_AF_LOC>(AFRes);
+  }
+
+  // SF
+  {
+    auto SignBitConst = _Constant(GetSrcSize(Op) * 8 - 1);
+
+    auto LshrOp = _Lshr(Res, SignBitConst);
+    SetRFLAG<FEXCore::X86State::RFLAG_SF_LOC>(LshrOp);
+  }
+
+  // PF
+  if (!CTX->Config.ABINoPF) {
+    auto EightBitMask = _Constant(0xFF);
+    auto PopCountOp = _Popcount(_And(Res, EightBitMask));
+    auto XorOp = _Xor(PopCountOp, _Constant(1));
+    SetRFLAG<FEXCore::X86State::RFLAG_PF_LOC>(XorOp);
+  } else {
+    _InvalidateFlags(1UL << FEXCore::X86State::RFLAG_PF_LOC);
+  }
+
+  // ZF
+  {
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ,
+        Res, _Constant(0), _Constant(1), _Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_ZF_LOC>(SelectOp);
+  }
+  // CF
+  if (UpdateCF) {
+    auto SelectOp = _Select(FEXCore::IR::COND_ULT, Res, Src2, _Constant(1), _Constant(0));
+
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(SelectOp);
+  }
+
+  // OF
+  {
+    auto NegOne = _Constant(~0ULL);
+    auto XorOp1 = _Xor(_Xor(Src1, Src2), NegOne);
+    auto XorOp2 = _Xor(Res, Src1);
+
+    OrderedNode *AndOp1 = _And(XorOp1, XorOp2);
+
+    switch (GetSrcSize(Op)) {
+    case 1:
+      AndOp1 = _Bfe(1, 7, AndOp1);
+    break;
+    case 2:
+      AndOp1 = _Bfe(1, 15, AndOp1);
+    break;
+    case 4:
+      AndOp1 = _Bfe(1, 31, AndOp1);
+    break;
+    case 8:
+      AndOp1 = _Bfe(1, 63, AndOp1);
+    break;
+    default: LOGMAN_MSG_A("Unknown BFESize: %d", GetSrcSize(Op)); break;
+    }
+    SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(AndOp1);
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_MUL(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *High) {
+  // PF/AF/ZF/SF
+  // Undefined
+  {
+    SetRFLAG<FEXCore::X86State::RFLAG_PF_LOC>(_Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_AF_LOC>(_Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_ZF_LOC>(_Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_SF_LOC>(_Constant(0));
+  }
+
+  // CF/OF
+  {
+    // CF and OF are set if the result of the operation can't be fit in to the destination register
+    // If the value can fit then the top bits will be zero
+
+    auto SignBit = _Sbfe(1, GetSrcSize(Op) * 8 - 1, Res);
+
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ, High, SignBit, _Constant(0), _Constant(1));
+
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(SelectOp);
+    SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(SelectOp);
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_UMUL(FEXCore::X86Tables::DecodedOp Op, OrderedNode *High) {
+  // AF/SF/PF/ZF
+  // Undefined
+  {
+    SetRFLAG<FEXCore::X86State::RFLAG_AF_LOC>(_Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_SF_LOC>(_Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_PF_LOC>(_Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_ZF_LOC>(_Constant(0));
+  }
+
+  // CF/OF
+  {
+    // CF and OF are set if the result of the operation can't be fit in to the destination register
+    // The result register will be all zero if it can't fit due to how multiplication behaves
+
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ, High, _Constant(0), _Constant(0), _Constant(1));
+
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(SelectOp);
+    SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(SelectOp);
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_Logical(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2) {
+  // AF
+  {
+    // Undefined
+    // Set to zero anyway
+    SetRFLAG<FEXCore::X86State::RFLAG_AF_LOC>(_Constant(0));
+  }
+
+  // SF
+  {
+    auto SignBitConst = _Constant(GetSrcSize(Op) * 8 - 1);
+
+    auto LshrOp = _Lshr(Res, SignBitConst);
+    SetRFLAG<FEXCore::X86State::RFLAG_SF_LOC>(LshrOp);
+  }
+
+  // PF
+  if (!CTX->Config.ABINoPF) {
+    auto EightBitMask = _Constant(0xFF);
+    auto PopCountOp = _Popcount(_And(Res, EightBitMask));
+    auto XorOp = _Xor(PopCountOp, _Constant(1));
+    SetRFLAG<FEXCore::X86State::RFLAG_PF_LOC>(XorOp);
+  } else {
+    _InvalidateFlags(1UL << FEXCore::X86State::RFLAG_PF_LOC);
+  }
+
+  // ZF
+  {
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ,
+        Res, _Constant(0), _Constant(1), _Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_ZF_LOC>(SelectOp);
+  }
+
+  // CF/OF
+  {
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(_Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(_Constant(0));
+  }
+}
+
+#define COND_FLAG_SET(cond, flag, newflag) \
+auto oldflag = GetRFLAG(FEXCore::X86State::flag);\
+auto newval = _Select(FEXCore::IR::COND_EQ, cond, _Constant(0), oldflag, newflag);\
+SetRFLAG<FEXCore::X86State::flag>(newval);
+
+void OpDispatchBuilder::GenerateFlags_ShiftLeft(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2) {
+  // CF
+  {
+    // Extract the last bit shifted in to CF
+    auto Size = _Constant(GetSrcSize(Op) * 8);
+    auto ShiftAmt = _Sub(Size, Src2);
+    auto LastBit = _And(_Lshr(Src1, ShiftAmt), _Constant(1));
+    COND_FLAG_SET(Src2, RFLAG_CF_LOC, LastBit);
+  }
+
+  // PF
+  if (!CTX->Config.ABINoPF) {
+    auto EightBitMask = _Constant(0xFF);
+    auto PopCountOp = _Popcount(_And(Res, EightBitMask));
+    auto XorOp = _Xor(PopCountOp, _Constant(1));
+    COND_FLAG_SET(Src2, RFLAG_PF_LOC, XorOp);
+  } else {
+    _InvalidateFlags(1UL << FEXCore::X86State::RFLAG_PF_LOC);
+  }
+
+  // AF
+  {
+    // Undefined
+    // Set to zero anyway
+    COND_FLAG_SET(Src2, RFLAG_AF_LOC, _Constant(0));
+  }
+
+  // ZF
+  {
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ,
+        Res, _Constant(0), _Constant(1), _Constant(0));
+    COND_FLAG_SET(Src2, RFLAG_ZF_LOC, SelectOp);
+  }
+
+  // SF
+  {
+    auto val = _Bfe(1, GetSrcSize(Op) * 8 - 1, Res);
+    COND_FLAG_SET(Src2, RFLAG_SF_LOC, val);
+  }
+
+  // OF
+  {
+    // In the case of left shift. OF is only set from the result of <Top Source Bit> XOR <Top Result Bit>
+    // When Shift > 1 then OF is undefined
+    auto val = _Bfe(1, GetSrcSize(Op) * 8 - 1, _Xor(Src1, Res));
+    COND_FLAG_SET(Src2, RFLAG_OF_LOC, val);
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_ShiftRight(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2) {
+  // CF
+  {
+    // Extract the last bit shifted in to CF
+    auto ShiftAmt = _Sub(Src2, _Constant(1));
+    auto LastBit = _And(_Lshr(Src1, ShiftAmt), _Constant(1));
+    COND_FLAG_SET(Src2, RFLAG_CF_LOC, LastBit);
+  }
+
+  // PF
+  if (!CTX->Config.ABINoPF) {
+    auto EightBitMask = _Constant(0xFF);
+    auto PopCountOp = _Popcount(_And(Res, EightBitMask));
+    auto XorOp = _Xor(PopCountOp, _Constant(1));
+    COND_FLAG_SET(Src2, RFLAG_PF_LOC, XorOp);
+  } else {
+    _InvalidateFlags(1UL << FEXCore::X86State::RFLAG_PF_LOC);
+  }
+
+  // AF
+  {
+    // Undefined
+    // Set to zero anyway
+    COND_FLAG_SET(Src2, RFLAG_AF_LOC, _Constant(0));
+  }
+
+  // ZF
+  {
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ,
+        Res, _Constant(0), _Constant(1), _Constant(0));
+    COND_FLAG_SET(Src2, RFLAG_ZF_LOC, SelectOp);
+  }
+
+  // SF
+  {
+    auto val =_Bfe(1, GetSrcSize(Op) * 8 - 1, Res);
+    COND_FLAG_SET(Src2, RFLAG_SF_LOC, val);
+  }
+
+  // OF
+  {
+    // Only defined when Shift is 1 else undefined
+    // OF flag is set if a sign change occurred
+    auto val = _Bfe(1, GetSrcSize(Op) * 8 - 1, _Xor(Src1, Res));
+    COND_FLAG_SET(Src2, RFLAG_OF_LOC, val);
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_SignShiftRight(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2) {
+  // CF
+  {
+    // Extract the last bit shifted in to CF
+    auto ShiftAmt = _Sub(Src2, _Constant(1));
+    auto LastBit = _And(_Lshr(Src1, ShiftAmt), _Constant(1));
+    COND_FLAG_SET(Src2, RFLAG_CF_LOC, LastBit);
+  }
+
+  // PF
+  if (!CTX->Config.ABINoPF) {
+    auto EightBitMask = _Constant(0xFF);
+    auto PopCountOp = _Popcount(_And(Res, EightBitMask));
+    auto XorOp = _Xor(PopCountOp, _Constant(1));
+    COND_FLAG_SET(Src2, RFLAG_PF_LOC, XorOp);
+  } else {
+    _InvalidateFlags(1UL << FEXCore::X86State::RFLAG_PF_LOC);
+  }
+
+  // AF
+  {
+    // Undefined
+    // Set to zero anyway
+    COND_FLAG_SET(Src2, RFLAG_AF_LOC, _Constant(0));
+  }
+
+  // ZF
+  {
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ,
+        Res, _Constant(0), _Constant(1), _Constant(0));
+    COND_FLAG_SET(Src2, RFLAG_ZF_LOC, SelectOp);
+  }
+
+  // SF
+  {
+    auto SignBitConst = _Constant(GetSrcSize(Op) * 8 - 1);
+
+    auto LshrOp = _Lshr(Res, SignBitConst);
+    COND_FLAG_SET(Src2, RFLAG_SF_LOC, LshrOp);
+  }
+
+  // OF
+  {
+    COND_FLAG_SET(Src2, RFLAG_OF_LOC, _Constant(0));
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_ShiftLeftImmediate(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, uint64_t Shift) {
+  // No flags changed if shift is zero
+  if (Shift == 0) return;
+
+  // CF
+  {
+    // Extract the last bit shifted in to CF
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(_Bfe(1, GetSrcSize(Op) * 8 - Shift, Src1));
+  }
+
+  // PF
+  if (!CTX->Config.ABINoPF) {
+    auto EightBitMask = _Constant(0xFF);
+    auto PopCountOp = _Popcount(_And(Res, EightBitMask));
+    auto XorOp = _Xor(PopCountOp, _Constant(1));
+    SetRFLAG<FEXCore::X86State::RFLAG_PF_LOC>(XorOp);
+  } else {
+    _InvalidateFlags(1UL << FEXCore::X86State::RFLAG_PF_LOC);
+  }
+
+  // AF
+  {
+    // Undefined
+    // Set to zero anyway
+    SetRFLAG<FEXCore::X86State::RFLAG_AF_LOC>(_Constant(0));
+  }
+
+  // ZF
+  {
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ,
+        Res, _Constant(0), _Constant(1), _Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_ZF_LOC>(SelectOp);
+  }
+
+  // SF
+  {
+    auto LshrOp = _Bfe(1, GetSrcSize(Op) * 8 - 1, Res);
+
+    SetRFLAG<FEXCore::X86State::RFLAG_SF_LOC>(LshrOp);
+
+    // OF
+    // In the case of left shift. OF is only set from the result of <Top Source Bit> XOR <Top Result Bit>
+    if (Shift == 1) {
+      auto SourceBit = _Bfe(1, GetSrcSize(Op) * 8 - 1, Src1);
+      SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(_Xor(SourceBit, LshrOp));
+    }
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_SignShiftRightImmediate(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, uint64_t Shift) {
+  // No flags changed if shift is zero
+  if (Shift == 0) return;
+
+  // CF
+  {
+    // Extract the last bit shifted in to CF
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(_Bfe(1, Shift-1, Src1));
+  }
+
+  // PF
+  if (!CTX->Config.ABINoPF) {
+    auto EightBitMask = _Constant(0xFF);
+    auto PopCountOp = _Popcount(_And(Res, EightBitMask));
+    auto XorOp = _Xor(PopCountOp, _Constant(1));
+    SetRFLAG<FEXCore::X86State::RFLAG_PF_LOC>(XorOp);
+  } else {
+    _InvalidateFlags(1UL << FEXCore::X86State::RFLAG_PF_LOC);
+  }
+
+  // AF
+  {
+    // Undefined
+    // Set to zero anyway
+    SetRFLAG<FEXCore::X86State::RFLAG_AF_LOC>(_Constant(0));
+  }
+
+  // ZF
+  {
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ,
+        Res, _Constant(0), _Constant(1), _Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_ZF_LOC>(SelectOp);
+  }
+
+  // SF
+  {
+    auto SignBitConst = _Constant(GetSrcSize(Op) * 8 - 1);
+
+    auto LshrOp = _Lshr(Res, SignBitConst);
+    SetRFLAG<FEXCore::X86State::RFLAG_SF_LOC>(LshrOp);
+
+    // OF
+    // Only defined when Shift is 1 else undefined
+    // Only is set if the top bit was set to 1 when shifted
+    // So it is set to same value as SF
+    if (Shift == 1) {
+      SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(_Constant(0));
+    }
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_ShiftRightImmediate(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, uint64_t Shift) {
+  // No flags changed if shift is zero
+  if (Shift == 0) return;
+
+  // CF
+  {
+    // Extract the last bit shifted in to CF
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(_Bfe(1, Shift-1, Src1));
+  }
+
+  // PF
+  if (!CTX->Config.ABINoPF) {
+    auto EightBitMask = _Constant(0xFF);
+    auto PopCountOp = _Popcount(_And(Res, EightBitMask));
+    auto XorOp = _Xor(PopCountOp, _Constant(1));
+    SetRFLAG<FEXCore::X86State::RFLAG_PF_LOC>(XorOp);
+  } else {
+    _InvalidateFlags(1UL << FEXCore::X86State::RFLAG_PF_LOC);
+  }
+
+  // AF
+  {
+    // Undefined
+    // Set to zero anyway
+    SetRFLAG<FEXCore::X86State::RFLAG_AF_LOC>(_Constant(0));
+  }
+
+  // ZF
+  {
+    auto SelectOp = _Select(FEXCore::IR::COND_EQ,
+        Res, _Constant(0), _Constant(1), _Constant(0));
+    SetRFLAG<FEXCore::X86State::RFLAG_ZF_LOC>(SelectOp);
+  }
+
+  // SF
+  {
+    auto SignBitConst = _Constant(GetSrcSize(Op) * 8 - 1);
+
+    auto LshrOp = _Lshr(Res, SignBitConst);
+    SetRFLAG<FEXCore::X86State::RFLAG_SF_LOC>(LshrOp);
+  }
+
+  // OF
+  {
+    // Only defined when Shift is 1 else undefined
+    // Is set to the MSB of the original value
+    if (Shift == 1) {
+      SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(_Bfe(1, GetSrcSize(Op) * 8 - 1, Src1));
+    }
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_RotateRight(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2) {
+  auto OpSize = GetSrcSize(Op) * 8;
+
+  // Extract the last bit shifted in to CF
+  auto NewCF = _Bfe(1, OpSize - 1, Res);
+
+  // CF
+  {
+    auto OldCF = GetRFLAG(FEXCore::X86State::RFLAG_CF_LOC);
+    auto CF = _Select(FEXCore::IR::COND_EQ, Src2, _Constant(0), OldCF, NewCF);
+
+    // Extract the last bit shifted in to CF
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(CF);
+  }
+
+  // OF
+  {
+    auto OldOF = GetRFLAG(FEXCore::X86State::RFLAG_OF_LOC);
+
+    // OF is set to the XOR of the new CF bit and the most significant bit of the result
+    auto NewOF = _Xor(_Bfe(1, OpSize - 2, Res), NewCF);
+
+    // If shift == 0, don't update flags
+    auto OF = _Select(FEXCore::IR::COND_EQ, Src2, _Constant(0), OldOF, NewOF);
+
+    SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(OF);
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_RotateLeft(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2) {
+  auto OpSize = GetSrcSize(Op) * 8;
+
+  // Extract the last bit shifted in to CF
+  //auto Size = _Constant(GetSrcSize(Res) * 8);
+  //auto ShiftAmt = _Sub(Size, Src2);
+  auto NewCF = _Bfe(1, 0, Res);
+
+  // CF
+  {
+    auto OldCF = GetRFLAG(FEXCore::X86State::RFLAG_CF_LOC);
+    auto CF = _Select(FEXCore::IR::COND_EQ, Src2, _Constant(0), OldCF, NewCF);
+
+    // Extract the last bit shifted in to CF
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(CF);
+  }
+
+  // OF
+  {
+    auto OldOF = GetRFLAG(FEXCore::X86State::RFLAG_OF_LOC);
+    // OF is set to the XOR of the new CF bit and the most significant bit of the result
+    auto NewOF = _Xor(_Bfe(1, OpSize - 1, Res), NewCF);
+
+    auto OF = _Select(FEXCore::IR::COND_EQ, Src2, _Constant(0), OldOF, NewOF);
+
+    // If shift == 0, don't update flags
+    SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(OF);
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_RotateRightImmediate(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, uint64_t Shift) {
+  if (Shift == 0) return;
+
+  auto OpSize = GetSrcSize(Op) * 8;
+
+  auto NewCF = _Bfe(1, OpSize - Shift, Src1);
+
+  // CF
+  {
+    // Extract the last bit shifted in to CF
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(NewCF);
+  }
+
+  // OF
+  {
+    if (Shift == 1) {
+      // OF is set to the XOR of the new CF bit and the most significant bit of the result
+      SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(_Xor(_Bfe(1, OpSize - 1, Res), NewCF));
+    }
+  }
+}
+
+void OpDispatchBuilder::GenerateFlags_RotateLeftImmediate(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, uint64_t Shift) {
+  if (Shift == 0) return;
+
+  auto OpSize = GetSrcSize(Op) * 8;
+
+  // CF
+  {
+    // Extract the last bit shifted in to CF
+    SetRFLAG<FEXCore::X86State::RFLAG_CF_LOC>(_Bfe(1, Shift, Src1));
+  }
+
+  // OF
+  {
+    if (Shift == 1) {
+      // OF is the top two MSBs XOR'd together
+      SetRFLAG<FEXCore::X86State::RFLAG_OF_LOC>(_Xor(_Bfe(1, OpSize - 1, Src1), _Bfe(1, OpSize - 2, Src1)));
+    }
+  }
+}
+
+}

External/FEXCore/Source/Interface/Core/SignalDelegator.cpp

@@ -0,0 +1,113 @@
+#include <FEXCore/Core/SignalDelegator.h>
+#include <FEXCore/Utils/LogManager.h>
+
+#include <unistd.h>
+#include <signal.h>
+
+namespace FEXCore {
+  struct ThreadState {
+    FEXCore::Core::InternalThreadState *Thread{};
+  };
+
+  thread_local ThreadState ThreadData{};
+
+  static bool IsSynchronous(int Signal) {
+    switch (Signal) {
+    case SIGBUS:
+    case SIGFPE:
+    case SIGILL:
+    case SIGSEGV:
+    case SIGTRAP:
+      return true;
+    default: break;
+    };
+    return false;
+  }
+
+  /**
+   * @brief Masks signals from the signal mask
+   *
+   * @param how Argument to sigmask. SIG_{BLOCK, SETMASK, UNBLOCK}
+   * @param Signal Which signal to set or -1 to sweep through them all
+   */
+  static void MaskSignals(int how, int Signal = -1) {
+    // If we have a helper thread, we need to mask a significant amount of signals so the an errant thread doesn't receive a signal that it shouldn't
+    sigset_t SignalSet{};
+    sigemptyset(&SignalSet);
+
+    if (Signal == -1) {
+      for (int i = 0; i <= SignalDelegator::MAX_SIGNALS; ++i) {
+        // If it is a synchronous signal then don't ignore it
+        if (IsSynchronous(i)) {
+          continue;
+        }
+
+        // Add this signal to the ignore list
+        sigaddset(&SignalSet, i);
+      }
+    }
+    else {
+      sigaddset(&SignalSet, Signal);
+    }
+
+    // Be warned, a thread will inherit the signal mask if created from this thread
+    int Result = pthread_sigmask(how, &SignalSet, nullptr);
+    if (Result != 0) {
+      LogMan::Msg::E("Couldn't register thread to mask signals");
+    }
+  }
+
+  void SignalDelegator::MaskThreadSignals() {
+    MaskSignals(SIG_BLOCK);
+  }
+
+  FEXCore::Core::InternalThreadState *SignalDelegator::GetTLSThread() {
+    return ThreadData.Thread;
+  }
+
+  void SignalDelegator::RegisterTLSState(FEXCore::Core::InternalThreadState *Thread) {
+    ThreadData.Thread = Thread;
+    RegisterFrontendTLSState(Thread);
+  }
+
+  void SignalDelegator::UninstallTLSState(FEXCore::Core::InternalThreadState *Thread) {
+    UninstallFrontendTLSState(Thread);
+    ThreadData.Thread = nullptr;
+  }
+
+  void SignalDelegator::RegisterHostSignalHandler(int Signal, HostSignalDelegatorFunction Func, bool Required) {
+    SetHostSignalHandler(Signal, Func, Required);
+    FrontendRegisterHostSignalHandler(Signal, Func, Required);
+  }
+
+  void SignalDelegator::RegisterFrontendHostSignalHandler(int Signal, HostSignalDelegatorFunction Func, bool Required) {
+    SetFrontendHostSignalHandler(Signal, Func, Required);
+    FrontendRegisterFrontendHostSignalHandler(Signal, Func, Required);
+  }
+
+  void SignalDelegator::HandleSignal(int Signal, void *Info, void *UContext) {
+    // Let the host take first stab at handling the signal
+    auto Thread = GetTLSThread();
+    HostSignalHandler &Handler = HostHandlers[Signal];
+
+    if (!Thread) {
+      LogMan::Msg::E("[%d] Thread has received a signal and hasn't registered itself with the delegate! Programming error!", ::gettid());
+    }
+    else {
+      if (Handler.Handler &&
+          Handler.Handler(Thread, Signal, Info, UContext)) {
+        // If the host handler handled the fault then we can continue now
+        return;
+      }
+
+      if (Handler.FrontendHandler &&
+          Handler.FrontendHandler(Thread, Signal, Info, UContext)) {
+        return;
+      }
+
+      // Now let the frontend handle the signal
+      // It's clearly a guest signal and this ends up being an OS specific issue
+      HandleGuestSignal(Thread, Signal, Info, UContext);
+    }
+  }
+}

External/FEXCore/Source/Interface/Core/VSyscall/VSyscall.inc

@@ -0,0 +1,285 @@
+// This is the vsyscall page for x86_64 guest code
+// This was compiled with nasm with the following source then exported to binary
+
+//BITS 64;
+//
+//align 4096, db 0xcc
+//  ; __NR_gettimeofday
+//  mov rax, 96
+//  syscall
+//  ret
+//
+//align 1024, db 0xcc
+//  ; __NR_time
+//  mov rax, 201
+//  syscall
+//  ret
+//
+//align 1024, db 0xcc
+//  ; __NR_getcpu
+//  mov rax, 309
+//  syscall
+//  ret
+//
+//align 4096, db 0xcc
+
+// We only want one of these pages per FEX process
+// One page
+const static uint8_t VSyscallData[0x1000] = {
+	0xB8, 0x60, 0x00, 0x00, 0x00, 0x0F, 0x05, 0xC3, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xB8, 0xC9, 0x00, 0x00, 0x00, 0x0F, 0x05, 0xC3, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xB8, 0x35, 0x01, 0x00, 0x00, 0x0F, 0x05, 0xC3, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC,
+	0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC
+};

External/FEXCore/Source/Interface/Core/X86DebugInfo.cpp

@@ -114,8 +114,6 @@ void InstallDebugInfo() {
   GenerateDebugTable(PrimaryInstGroupOps, PrimaryGroupOpTable);
 
   GenerateDebugTable(SecondInstGroupOps, SecondaryExtensionOpTable);
-
-  LogMan::Msg::D("Installing debug info");
 }
 }
 #endif

External/FEXCore/Source/Interface/Core/X86HelperGen.cpp

@@ -6,12 +6,15 @@ $end_info$
 */
 
 #include "Interface/Core/X86HelperGen.h"
+
+#include <FEXCore/Config/Config.h>
 #include <FEXCore/Utils/Allocator.h>
 
+#include <cstdint>
 #include <cstring>
-#include <stdlib.h>
 #include <vector>
 #include <sys/mman.h>
+#include <bits/mman-map-flags-generic.h>
 
 namespace FEXCore {
 constexpr size_t CODE_SIZE = 0x1000;
@@ -50,14 +53,14 @@ void* X86GeneratedCode::AllocateGuestCodeSpace(size_t Size) {
   // We need to have the sigret handler in the lower 32bits of memory space
   // Scan top down and try to allocate a location
   for (size_t Location = 0xFFFF'E000; Location != 0x0; Location -= 0x1000) {
-    void *Ptr = FEXCore::Allocator::mmap(reinterpret_cast<void*>(Location), Size, PROT_READ | PROT_WRITE, MAP_FIXED_NOREPLACE | MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+    void *Ptr = ::mmap(reinterpret_cast<void*>(Location), Size, PROT_READ | PROT_WRITE, MAP_FIXED_NOREPLACE | MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 
     if (Ptr != MAP_FAILED &&
         reinterpret_cast<uintptr_t>(Ptr) >= LOCATION_MAX) {
       // Failed to map in the lower 32bits
       // Try again
       // Can happen in the case that host kernel ignores MAP_FIXED_NOREPLACE
-      FEXCore::Allocator::munmap(Ptr, Size);
+      ::munmap(Ptr, Size);
       continue;
     }
 

External/FEXCore/Source/Interface/Core/X86HelperGen.h

@@ -5,8 +5,8 @@ $end_info$
 */
 
 #pragma once
-#include <FEXCore/Config/Config.h>
 
+#include <stddef.h>
 #include <stdint.h>
 
 namespace FEXCore {

External/FEXCore/Source/Interface/Core/X86Tables.cpp

@@ -5,14 +5,8 @@ tags: frontend|x86-tables
 $end_info$
 */
 
-#include <FEXCore/Utils/LogManager.h>
-
 #include <FEXCore/Core/Context.h>
 #include <FEXCore/Debug/X86Tables.h>
-#include <array>
-#include <cstdint>
-#include <tuple>
-#include <vector>
 
 namespace FEXCore::X86Tables {
 
@@ -108,7 +102,6 @@ void InitializeInfoTables(Context::OperatingMode Mode) {
 
 #ifndef NDEBUG
   X86InstDebugInfo::InstallDebugInfo();
-  LogMan::Msg::D("X86Tables had %ld total insts, and %ld labeled as understood", Total, NumInsts);
 #endif
 }
 

External/FEXCore/Source/Interface/Core/X86Tables/BaseTables.cpp

@@ -7,6 +7,9 @@ $end_info$
 #include "Interface/Core/X86Tables/X86Tables.h"
 
 #include <FEXCore/Core/Context.h>
+#include <FEXCore/Debug/X86Tables.h>
+
+#include <iterator>
 
 namespace FEXCore::X86Tables {
 using namespace InstFlags;

External/FEXCore/Source/Interface/Core/X86Tables/DDDTables.cpp

@@ -6,6 +6,10 @@ $end_info$
 
 #include "Interface/Core/X86Tables/X86Tables.h"
 
+#include <FEXCore/Debug/X86Tables.h>
+
+#include <iterator>
+
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
 

External/FEXCore/Source/Interface/Core/X86Tables/EVEXTables.cpp

@@ -6,6 +6,10 @@ $end_info$
 
 #include "Interface/Core/X86Tables/X86Tables.h"
 
+#include <FEXCore/Debug/X86Tables.h>
+
+#include <iterator>
+
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
 

External/FEXCore/Source/Interface/Core/X86Tables/H0F38Tables.cpp

@@ -5,6 +5,10 @@ $end_info$
 */
 
 #include "Interface/Core/X86Tables/X86Tables.h"
+#include <FEXCore/Debug/X86Tables.h>
+
+#include <iterator>
+#include <stdint.h>
 
 namespace FEXCore::X86Tables {
 using namespace InstFlags;

External/FEXCore/Source/Interface/Core/X86Tables/H0F3ATables.cpp

@@ -6,6 +6,12 @@ $end_info$
 
 #include "Interface/Core/X86Tables/X86Tables.h"
 
+#include <FEXCore/Debug/X86Tables.h>
+#include <FEXCore/Core/Context.h>
+
+#include <iterator>
+#include <stdint.h>
+
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
 

External/FEXCore/Source/Interface/Core/X86Tables/PrimaryGroupTables.cpp

@@ -6,6 +6,11 @@ $end_info$
 
 #include "Interface/Core/X86Tables/X86Tables.h"
 
+#include <FEXCore/Core/Context.h>
+#include <FEXCore/Debug/X86Tables.h>
+
+#include <iterator>
+
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
 

External/FEXCore/Source/Interface/Core/X86Tables/SecondaryGroupTables.cpp

@@ -6,6 +6,11 @@ $end_info$
 
 #include "Interface/Core/X86Tables/X86Tables.h"
 
+#include <FEXCore/Debug/X86Tables.h>
+
+#include <iterator>
+#include <stdint.h>
+
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
 

External/FEXCore/Source/Interface/Core/X86Tables/SecondaryModRMTables.cpp

@@ -6,6 +6,10 @@ $end_info$
 
 #include "Interface/Core/X86Tables/X86Tables.h"
 
+#include <FEXCore/Debug/X86Tables.h>
+
+#include <iterator>
+
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
 

External/FEXCore/Source/Interface/Core/X86Tables/SecondaryTables.cpp

@@ -4,7 +4,12 @@ tags: frontend|x86-tables
 $end_info$
 */
 
- #include "Interface/Core/X86Tables/X86Tables.h"
+#include "Interface/Core/X86Tables/X86Tables.h"
+
+#include <FEXCore/Core/Context.h>
+#include <FEXCore/Debug/X86Tables.h>
+
+#include <iterator>
 
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
@@ -27,7 +32,7 @@ void InitializeSecondaryTables(Context::OperatingMode Mode) {
     {0x0B, 1, X86InstInfo{"UD2",        TYPE_INST, FLAGS_DEBUG | FLAGS_BLOCK_END | FLAGS_NO_OVERLAY,                                                    0, nullptr}},
     {0x0C, 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NO_OVERLAY,                                                                                 0, nullptr}},
     {0x0D, 1, X86InstInfo{"",           TYPE_GROUP_P, FLAGS_MODRM | FLAGS_NO_OVERLAY,                                                                                 0, nullptr}},
-    {0x0E, 1, X86InstInfo{"FEMMS",      TYPE_INST, FLAGS_BLOCK_END | FLAGS_NO_OVERLAY,                                                            0, nullptr}},
+    {0x0E, 1, X86InstInfo{"FEMMS",      TYPE_INST, FLAGS_NO_OVERLAY,                                                            0, nullptr}},
     {0x0F, 1, X86InstInfo{"",           TYPE_3DNOW_TABLE, FLAGS_NO_OVERLAY,                                                                             0, nullptr}},
 
     {0x10, 1, X86InstInfo{"MOVUPS",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS,                                                 0, nullptr}},

External/FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp

@@ -6,6 +6,9 @@ $end_info$
 
 #include "Interface/Core/X86Tables/X86Tables.h"
 
+#include <FEXCore/Debug/X86Tables.h>
+#include <iterator>
+
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
 

External/FEXCore/Source/Interface/Core/X86Tables/X86Tables.h

@@ -6,6 +6,7 @@ $end_info$
 
 #pragma once
 #include <FEXCore/Debug/X86Tables.h>
+#include <FEXCore/Core/Context.h>
 
 #include <FEXCore/Utils/LogManager.h>
 
@@ -113,5 +114,7 @@ static inline void GenerateX87Table(X86InstInfo *FinalTable, U16U8InfoStruct con
   }
 };
 
+void InitializeInfoTables(Context::OperatingMode Mode);
+
 }
 

External/FEXCore/Source/Interface/Core/X86Tables/X87Tables.cpp

@@ -6,6 +6,9 @@ $end_info$
 
 #include "Interface/Core/X86Tables/X86Tables.h"
 
+#include <FEXCore/Debug/X86Tables.h>
+#include <iterator>
+
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
 

External/FEXCore/Source/Interface/Core/X86Tables/XOPTables.cpp

@@ -6,6 +6,11 @@ $end_info$
 
 #include "Interface/Core/X86Tables/X86Tables.h"
 
+#include <FEXCore/Debug/X86Tables.h>
+
+#include <iterator>
+#include <stdint.h>
+
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
 

External/FEXCore/Source/Interface/HLE/Thunks/Thunks.cpp

@@ -5,21 +5,24 @@ tags: glue|thunks
 $end_info$
 */
 
+#include <FEXCore/Config/Config.h>
+#include <FEXCore/Core/CoreState.h>
+#include <FEXCore/Debug/InternalThreadState.h>
 #include <FEXCore/Utils/LogManager.h>
+#include <FEXCore/IR/IR.h>
 #include "Thunks.h"
 
-#include "stdio.h"
 #include <dlfcn.h>
 
-
-#include <string>
-#include <map>
-#include <array>
 #include <Interface/Context/Context.h>
-#include "Interface/Core/InternalThreadState.h"
 #include "FEXCore/Core/X86Enums.h"
-#include <mutex>
+#include <malloc.h>
+#include <map>
+#include <memory>
 #include <shared_mutex>
+#include <stdint.h>
+#include <string>
+#include <utility>
 
 struct LoadlibArgs {
     const char *Name;
@@ -30,7 +33,6 @@ static thread_local FEXCore::Core::InternalThreadState *Thread;
 
 
 namespace FEXCore {
-
     struct ExportEntry { uint8_t *sha256; ThunkedFunction* Fn; };
 
     class ThunkHandler_impl final: public ThunkHandler {
@@ -48,18 +50,17 @@ namespace FEXCore {
             Set arg0/1 to arg regs, use CTX::HandleCallback to handle the callback
         */
         static void CallCallback(void *callback, void *arg0, void* arg1) {
-            Thread->CurrentFrame->State.gregs[FEXCore::X86State::REG_RDI] = (uintptr_t)arg0;
-            Thread->CurrentFrame->State.gregs[FEXCore::X86State::REG_RSI] = (uintptr_t)arg1;
+          Thread->CurrentFrame->State.gregs[FEXCore::X86State::REG_RDI] = (uintptr_t)arg0;
+          Thread->CurrentFrame->State.gregs[FEXCore::X86State::REG_RSI] = (uintptr_t)arg1;
 
-            Thread->CTX->HandleCallback((uintptr_t)callback);
+          Thread->CTX->HandleCallback(Thread, (uintptr_t)callback);
         }
 
         static void LoadLib(void *ArgsV) {
+            auto CTX = Thread->CTX;
 
             auto Args = reinterpret_cast<LoadlibArgs*>(ArgsV);
 
-            auto CTX = Thread->CTX;
-
             auto Name = Args->Name;
             auto CallbackThunks = Args->CallbackThunks;
 
@@ -121,7 +122,9 @@ namespace FEXCore {
         }
 
         ThunkHandler_impl() {
+        }
 
+        ~ThunkHandler_impl() {
         }
     };
 

External/FEXCore/Source/Interface/HLE/Thunks/Thunks.h

@@ -5,12 +5,19 @@ $end_info$
 */
 
 #pragma once
-#include <FEXCore/IR/IR.h>
+
+namespace FEXCore::Context {
+  struct Context;
+}
 
 namespace FEXCore::Core {
   struct InternalThreadState;
 }
 
+namespace FEXCore::IR {
+  struct SHA256Sum;
+}
+
 namespace FEXCore {
     typedef void ThunkedFunction(void* ArgsRv);
 
@@ -22,4 +29,4 @@ namespace FEXCore {
 
         static ThunkHandler* Create();
     };
-};
+};

External/FEXCore/Source/Interface/IR/IR.json

@@ -605,10 +605,17 @@
     "FillRegister": {
       "Desc": ["Fills a register from a spill slot",
                "Spill slots are register allocated and has live ranges calculated to handle slot calculation",
-               "```diff\n- !Don't use this op. It is for RA to handle spilling and filling!\n```"
+               "```diff\n- !Don't use this op. It is for RA to handle spilling and filling!\n```",
+               "",
+               "The OriginalValue SSA arg points at the original SSA value spilled, and only exists for",
+               "RA validation purposes"
               ],
 
       "OpClass": "Memory",
+      "SSAArgs": "1",
+      "SSANames": [
+        "OriginalValue"
+      ],
       "HasDest": true,
       "DestClass": "Complex",
       "Args": [
@@ -1462,6 +1469,24 @@
       ]
     },
 
+    "AtomicFetchNeg": {
+      "HasSideEffects": true,
+      "Desc": ["Atomic integer fetch and two's complement negate",
+               "Dest is the value prior to operating on the value in memory"
+              ],
+      "OpClass": "Atomic",
+      "HasDest": true,
+      "DestClass": "GPR",
+      "DestSize": "Size",
+      "SSAArgs": "1",
+      "SSANames": [
+        "Addr"
+      ],
+      "Args": [
+        "uint8_t", "Size"
+      ]
+    },
+
     "VExtractToGPR": {
       "Desc": ["Extracts an element from a vector and places it in a GPR",
                "The element that is extracted from the vector is zero extended to the GPR size"

External/FEXCore/Source/Interface/IR/IRDumper.cpp

@@ -7,9 +7,14 @@ $end_info$
 
 #include <FEXCore/IR/IR.h>
 #include <FEXCore/IR/IntrusiveIRList.h>
-#include <FEXCore/Utils/LogManager.h>
-#include "Interface/IR/Passes/RegisterAllocationPass.h"
+#include <FEXCore/IR/RegisterAllocationData.h>
 
+#include <algorithm>
+#include <array>
+#include <ostream>
+#include <stdint.h>
+#include <string>
+#include <string_view>
 #include  <iomanip>
 
 namespace FEXCore::IR {

External/FEXCore/Source/Interface/IR/IREmitter.cpp

@@ -5,7 +5,15 @@ tags: ir|emitter
 $end_info$
 */
 
+#include <FEXCore/IR/IR.h>
 #include <FEXCore/IR/IREmitter.h>
+#include <FEXCore/IR/IntrusiveIRList.h>
+#include <FEXCore/Utils/LogManager.h>
+
+#include <array>
+#include <stdint.h>
+#include <string.h>
+#include <vector>
 
 namespace FEXCore::IR {
 void IREmitter::ResetWorkingList() {
@@ -18,12 +26,12 @@ void IREmitter::ResetWorkingList() {
   CurrentCodeBlock = nullptr;
 }
 
-void IREmitter::ReplaceAllUsesWithRange(OrderedNode *Node, OrderedNode *NewNode, AllNodesIterator After, AllNodesIterator End) {
+void IREmitter::ReplaceAllUsesWithRange(OrderedNode *Node, OrderedNode *NewNode, AllNodesIterator Begin, AllNodesIterator End) {
   uintptr_t ListBegin = DualListData.ListBegin();
   auto NodeId = Node->Wrapped(ListBegin).ID();
 
-  while (After != End) {
-    auto [RealNode, IROp] = After();
+  while (Begin != End) {
+    auto [RealNode, IROp] = Begin();
 
     uint8_t NumArgs = IR::GetArgs(IROp->Op);
     for (uint8_t i = 0; i < NumArgs; ++i) {
@@ -39,7 +47,7 @@ void IREmitter::ReplaceAllUsesWithRange(OrderedNode *Node, OrderedNode *NewNode,
       }
     }
 
-    ++After;
+    ++Begin;
   }
 }
 
@@ -82,7 +90,7 @@ IREmitter::IRPair<IROp_CodeBlock> IREmitter::CreateNewCodeBlockAfter(OrderedNode
   if (insertAfter) {
     LinkCodeBlocks(insertAfter, CodeNode);
   } else {
-    LOGMAN_THROW_A(CurrentCodeBlock != nullptr, "CurrentCodeBlock must not be null here");
+    LOGMAN_THROW_A_FMT(CurrentCodeBlock != nullptr, "CurrentCodeBlock must not be null here");
 
     // Find last block
     auto LastBlock = CurrentCodeBlock;
@@ -101,7 +109,7 @@ IREmitter::IRPair<IROp_CodeBlock> IREmitter::CreateNewCodeBlockAfter(OrderedNode
 
 void IREmitter::SetCurrentCodeBlock(OrderedNode *Node) {
   CurrentCodeBlock = Node;
-  LOGMAN_THROW_A(Node->Op(DualListData.DataBegin())->Op == OP_CODEBLOCK, "Node wasn't codeblock. It was '%s'", std::string(IR::GetName(Node->Op(DualListData.DataBegin())->Op)).c_str());
+  LOGMAN_THROW_A_FMT(Node->Op(DualListData.DataBegin())->Op == OP_CODEBLOCK, "Node wasn't codeblock. It was '{}'", IR::GetName(Node->Op(DualListData.DataBegin())->Op));
   SetWriteCursor(Node->Op(DualListData.DataBegin())->CW<IROp_CodeBlock>()->Begin.GetNode(DualListData.ListBegin()));
 }
 

External/FEXCore/Source/Interface/IR/IRParser.cpp

@@ -5,16 +5,24 @@ tags: ir|parser
 $end_info$
 */
 
-#include <string>
-#include <vector>
-#include <istream>
-#include <unordered_map>
-
 #include <FEXCore/IR/IR.h>
 #include <FEXCore/IR/IntrusiveIRList.h>
 #include <FEXCore/IR/IREmitter.h>
+#include <FEXCore/Utils/LogManager.h>
 
-
+#include <algorithm>
+#include <array>
+#include <cstdint>
+#include <errno.h>
+#include <memory>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string>
+#include <string_view>
+#include <utility>
+#include <vector>
+#include <istream>
+#include <unordered_map>
 
 namespace FEXCore::IR {
 namespace {
@@ -70,8 +78,6 @@ std::string DecodeErrorToString(DecodeFailure Failure) {
   return "Unknown Error";
 }
 
-std::unordered_map<std::string_view, FEXCore::IR::IROps> NameToOpMap;
-
 class IRParser: public FEXCore::IR::IREmitter {
   public:
   template<typename Type>
@@ -299,9 +305,10 @@ class IRParser: public FEXCore::IR::IREmitter {
   std::unordered_map<std::string, OrderedNode*> SSANameMapper;
   std::vector<LineDefinition> Defs;
   LineDefinition *CurrentDef{};
+  std::unordered_map<std::string_view, FEXCore::IR::IROps> NameToOpMap;
 
   IRParser(std::istream *text) {
-    InitializeStaticTables();
+    InitializeNameMap();
 
     std::string TmpLine;
     while (!text->eof()) {
@@ -630,7 +637,7 @@ class IRParser: public FEXCore::IR::IREmitter {
 		return true;
 	}
 
-  void InitializeStaticTables() {
+  void InitializeNameMap() {
     if (NameToOpMap.empty()) {
       for (FEXCore::IR::IROps Op = FEXCore::IR::IROps::OP_DUMMY;
          Op <= FEXCore::IR::IROps::OP_LAST;