[a64] Implement an ARM64 backend

.gitmodules

@@ -85,3 +85,6 @@
 [submodule "third_party/VulkanMemoryAllocator"]
 	path = third_party/VulkanMemoryAllocator
 	url = https://github.com/GPUOpen-LibrariesAndSDKs/VulkanMemoryAllocator.git
+[submodule "third_party/oaknut"]
+	path = third_party/oaknut
+	url = https://github.com/merryhime/oaknut.git

premake5.lua

@@ -54,7 +54,7 @@ filter("configurations:Checked")
   defines({
     "DEBUG",
   })
-filter({"configurations:Checked", "platforms:Windows"})
+filter({"configurations:Checked", "platforms:Windows-*"})
   buildoptions({
     "/RTCsu",           -- Full Run-Time Checks.
   })
@@ -153,7 +153,7 @@ filter("platforms:Android-*")
     "log",
   })
 
-filter("platforms:Windows")
+filter("platforms:Windows-*")
   system("windows")
   toolset("msc")
   buildoptions({
@@ -179,8 +179,12 @@ filter("platforms:Windows")
     "_CRT_SECURE_NO_WARNINGS",
     "WIN32",
     "_WIN64=1",
-    "_AMD64=1",
   })
+  filter("architecture:x86_64")
+    defines({
+      "_AMD64=1",
+    })
+  filter({})
   linkoptions({
     "/ignore:4006",  -- Ignores complaints about empty obj files.
     "/ignore:4221",
@@ -198,7 +202,7 @@ filter("platforms:Windows")
   })
 
 -- Embed the manifest for things like dependencies and DPI awareness.
-filter({"platforms:Windows", "kind:ConsoleApp or WindowedApp"})
+filter({"platforms:Windows-*", "kind:ConsoleApp or WindowedApp"})
   files({
     "src/xenia/base/app_win32.manifest"
   })
@@ -228,7 +232,12 @@ workspace("xenia")
         ["ARCHS"] = "x86_64"
       })
     elseif os.istarget("windows") then
-      platforms({"Windows"})
+      platforms({"Windows-ARM64", "Windows-x86_64"})
+      filter("platforms:Windows-ARM64")
+        architecture("ARM64")
+      filter("platforms:Windows-x86_64")
+        architecture("x86_64")
+      filter({})
       -- 10.0.15063.0: ID3D12GraphicsCommandList1::SetSamplePositions.
       -- 10.0.19041.0: D3D12_HEAP_FLAG_CREATE_NOT_ZEROED.
       -- 10.0.22000.0: DWMWA_WINDOW_CORNER_PREFERENCE.
@@ -284,7 +293,13 @@ workspace("xenia")
   include("src/xenia/apu/nop")
   include("src/xenia/base")
   include("src/xenia/cpu")
-  include("src/xenia/cpu/backend/x64")
+
+  filter("architecture:x86_64")
+    include("src/xenia/cpu/backend/x64")
+  filter("architecture:ARM64")
+    include("src/xenia/cpu/backend/a64")
+  filter({})
+
   include("src/xenia/debug/ui")
   include("src/xenia/gpu")
   include("src/xenia/gpu/null")

src/xenia/app/premake5.lua

@@ -32,6 +32,7 @@ project("xenia-app")
     "libavcodec",
     "libavutil",
     "mspack",
+    "SDL2",
     "snappy",
     "xxhash",
   })
@@ -72,13 +73,18 @@ project("xenia-app")
       "xenia-cpu-backend-x64",
     })
 
+  filter("architecture:ARM64")
+    links({
+      "xenia-cpu-backend-a64",
+    })
+
   -- TODO(Triang3l): The emulator itself on Android.
   filter("platforms:not Android-*")
     files({
       "xenia_main.cc",
     })
 
-  filter("platforms:Windows")
+  filter("platforms:Windows-*")
     files({
       "main_resources.rc",
     })
@@ -104,7 +110,7 @@ project("xenia-app")
       "SDL2",
     })
 
-  filter("platforms:Windows")
+  filter("platforms:Windows-*")
     links({
       "xenia-apu-xaudio2",
       "xenia-gpu-d3d12",
@@ -113,13 +119,13 @@ project("xenia-app")
       "xenia-ui-d3d12",
     })
 
-  filter({"platforms:Windows", SINGLE_LIBRARY_FILTER})
+  filter({"platforms:Windows-*", SINGLE_LIBRARY_FILTER})
     links({
       "xenia-gpu-d3d12-trace-viewer",
       "xenia-ui-window-d3d12-demo",
     })
 
-  filter("platforms:Windows")
+  filter("platforms:Windows-*")
     -- Only create the .user file if it doesn't already exist.
     local user_file = project_root.."/build/xenia-app.vcxproj.user"
     if not os.isfile(user_file) then

src/xenia/base/clock.cc

@@ -21,8 +21,9 @@ DEFINE_bool(clock_no_scaling, false,
             "Guest system time is directly pulled from host.",
             "CPU");
 DEFINE_bool(clock_source_raw, false,
-            "Use the RDTSC instruction as the time source. "
-            "Host CPU must support invariant TSC.",
+            "On x64, Use the RDTSC instruction as the time source. Requires "
+            "invariant TSC. "
+            "On a64, Use the CNTVCT_EL0 register as the time source",
             "CPU");
 
 namespace xe {

src/xenia/base/clock.h

@@ -18,6 +18,8 @@
 
 #if XE_ARCH_AMD64
 #define XE_CLOCK_RAW_AVAILABLE 1
+#elif XE_ARCH_ARM64
+#define XE_CLOCK_RAW_AVAILABLE 1
 #endif
 
 DECLARE_bool(clock_no_scaling);

src/xenia/base/clock_a64.cc

@@ -0,0 +1,50 @@
+/**
+ ******************************************************************************
+ * Xenia : Xbox 360 Emulator Research Project                                 *
+ ******************************************************************************
+ * Copyright 2024 Ben Vanik. All rights reserved.                             *
+ * Released under the BSD license - see LICENSE in the root for more details. *
+ ******************************************************************************
+ */
+
+#include "xenia/base/clock.h"
+#include "xenia/base/platform.h"
+
+#if XE_ARCH_ARM64 && XE_CLOCK_RAW_AVAILABLE
+
+#include "xenia/base/logging.h"
+
+#ifdef _MSC_VER
+#include <arm64_neon.h>
+#include <intrin.h>
+#else
+#include <arm_neon.h>
+#endif
+
+// Wrap all these different cpu compiler intrinsics.
+#if XE_COMPILER_MSVC
+constexpr int32_t CNTFRQ_EL0 = ARM64_SYSREG(3, 3, 14, 0, 0);
+constexpr int32_t CNTVCT_EL0 = ARM64_SYSREG(3, 3, 14, 0, 2);
+#define xe_cpu_mrs(reg) _ReadStatusReg(reg)
+#elif XE_COMPILER_CLANG || XE_COMPILER_GNUC
+constexpr int32_t CNTFRQ_EL0 = 0b11'011'1110'0000'000;
+constexpr int32_t CNTVCT_EL0 = 0b11'011'1110'0000'010;
+
+uint64_t xe_cpu_mrs(uint32_t reg) {
+  uint64_t result;
+  __asm__ volatile("mrs \t%0," #reg : "=r"(result));
+  return result;
+}
+#else
+#error \
+    "No cpu instruction wrappers xe_cpu_mrs(CNTVCT_EL0); for current compiler implemented."
+#endif
+
+namespace xe {
+
+uint64_t Clock::host_tick_frequency_raw() { return xe_cpu_mrs(CNTFRQ_EL0); }
+uint64_t Clock::host_tick_count_raw() { return xe_cpu_mrs(CNTVCT_EL0); }
+
+}  // namespace xe
+
+#endif

src/xenia/base/exception_handler_win.cc

@@ -36,12 +36,22 @@ LONG CALLBACK ExceptionHandlerCallback(PEXCEPTION_POINTERS ex_info) {
   }
 
   HostThreadContext thread_context;
+
+#if XE_ARCH_AMD64
   thread_context.rip = ex_info->ContextRecord->Rip;
   thread_context.eflags = ex_info->ContextRecord->EFlags;
   std::memcpy(thread_context.int_registers, &ex_info->ContextRecord->Rax,
               sizeof(thread_context.int_registers));
   std::memcpy(thread_context.xmm_registers, &ex_info->ContextRecord->Xmm0,
               sizeof(thread_context.xmm_registers));
+#elif XE_ARCH_ARM64
+  thread_context.pc = ex_info->ContextRecord->Pc;
+  thread_context.cpsr = ex_info->ContextRecord->Cpsr;
+  std::memcpy(thread_context.x, &ex_info->ContextRecord->X,
+              sizeof(thread_context.x));
+  std::memcpy(thread_context.v, &ex_info->ContextRecord->V,
+              sizeof(thread_context.v));
+#endif
 
   // https://msdn.microsoft.com/en-us/library/ms679331(v=vs.85).aspx
   // https://msdn.microsoft.com/en-us/library/aa363082(v=vs.85).aspx
@@ -78,6 +88,7 @@ LONG CALLBACK ExceptionHandlerCallback(PEXCEPTION_POINTERS ex_info) {
   for (size_t i = 0; i < xe::countof(handlers_) && handlers_[i].first; ++i) {
     if (handlers_[i].first(&ex, handlers_[i].second)) {
       // Exception handled.
+#if XE_ARCH_AMD64
       ex_info->ContextRecord->Rip = thread_context.rip;
       ex_info->ContextRecord->EFlags = thread_context.eflags;
       uint32_t modified_register_index;
@@ -98,6 +109,28 @@ LONG CALLBACK ExceptionHandlerCallback(PEXCEPTION_POINTERS ex_info) {
                     &thread_context.xmm_registers[modified_register_index],
                     sizeof(vec128_t));
       }
+#elif XE_ARCH_ARM64
+      ex_info->ContextRecord->Pc = thread_context.pc;
+      ex_info->ContextRecord->Cpsr = thread_context.cpsr;
+      uint32_t modified_register_index;
+      uint16_t modified_int_registers_remaining = ex.modified_x_registers();
+      while (xe::bit_scan_forward(modified_int_registers_remaining,
+                                  &modified_register_index)) {
+        modified_int_registers_remaining &=
+            ~(UINT16_C(1) << modified_register_index);
+        ex_info->ContextRecord->X[modified_register_index] =
+            thread_context.x[modified_register_index];
+      }
+      uint16_t modified_xmm_registers_remaining = ex.modified_v_registers();
+      while (xe::bit_scan_forward(modified_xmm_registers_remaining,
+                                  &modified_register_index)) {
+        modified_xmm_registers_remaining &=
+            ~(UINT16_C(1) << modified_register_index);
+        std::memcpy(&ex_info->ContextRecord->V + modified_register_index,
+                    &thread_context.v[modified_register_index],
+                    sizeof(vec128_t));
+      }
+#endif
       return EXCEPTION_CONTINUE_EXECUTION;
     }
   }

src/xenia/base/host_thread_context.cc

@@ -67,7 +67,7 @@ std::string HostThreadContext::GetStringFromValue(HostRegister reg,
     case Arm64Register::kPc:
       return hex ? string_util::to_hex_string(pc) : std::to_string(pc);
     case Arm64Register::kPstate:
-      return hex ? string_util::to_hex_string(pstate) : std::to_string(pstate);
+      return hex ? string_util::to_hex_string(cpsr) : std::to_string(cpsr);
     case Arm64Register::kFpsr:
       return hex ? string_util::to_hex_string(fpsr) : std::to_string(fpsr);
     case Arm64Register::kFpcr:

src/xenia/base/host_thread_context.h

@@ -202,7 +202,7 @@ class HostThreadContext {
   uint64_t x[31];
   uint64_t sp;
   uint64_t pc;
-  uint64_t pstate;
+  uint32_t cpsr;
   uint32_t fpsr;
   uint32_t fpcr;
   vec128_t v[32];

src/xenia/base/main_init_win.cc

@@ -11,6 +11,8 @@
 
 #include <cstdlib>
 
+#if XE_ARCH_AMD64
+
 // Includes Windows headers, so it goes after platform_win.h.
 #include "third_party/xbyak/xbyak/xbyak_util.h"
 
@@ -39,3 +41,5 @@ class StartupAvxCheck {
 #pragma warning(suppress : 4073)
 #pragma init_seg(lib)
 static StartupAvxCheck gStartupAvxCheck;
+
+#endif

src/xenia/base/math.h

@@ -31,6 +31,8 @@
 
 #if XE_ARCH_AMD64
 #include <xmmintrin.h>
+#elif XE_ARCH_ARM64
+#include <arm64_neon.h>
 #endif
 
 namespace xe {
@@ -133,10 +135,17 @@ constexpr inline uint32_t bit_count(T v) {
 }
 #else
 #if XE_COMPILER_MSVC || XE_COMPILER_INTEL
+#if XE_ARCH_AMD64
 inline uint32_t bit_count(uint32_t v) { return __popcnt(v); }
 inline uint32_t bit_count(uint64_t v) {
   return static_cast<uint32_t>(__popcnt64(v));
 }
+#elif XE_ARCH_ARM64
+inline uint32_t bit_count(uint32_t v) { return _CountOneBits(v); }
+inline uint32_t bit_count(uint64_t v) {
+  return static_cast<uint32_t>(_CountOneBits64(v));
+}
+#endif
 #elif XE_COMPILER_GCC || XE_COMPILER_CLANG
 static_assert(sizeof(unsigned int) == sizeof(uint32_t));
 static_assert(sizeof(unsigned long long) == sizeof(uint64_t));
@@ -376,6 +385,24 @@ template <int N>
 int64_t m128_i64(const __m128& v) {
   return m128_i64<N>(_mm_castps_pd(v));
 }
+#elif XE_ARCH_ARM64
+// Utilities for NEON values.
+template <int N>
+float m128_f32(const float32x4_t& v) {
+  return vgetq_lane_f32(v, N);
+}
+template <int N>
+int32_t m128_i32(const int32x4_t& v) {
+  return vgetq_lane_s32(v, N);
+}
+template <int N>
+double m128_f64(const float64x2_t& v) {
+  return vgetq_lane_f64(v, N);
+}
+template <int N>
+int64_t m128_i64(const int64x2_t& v) {
+  return vgetq_lane_s64(v, N);
+}
 #endif
 
 // Similar to the C++ implementation of XMConvertFloatToHalf and

src/xenia/base/platform.h

@@ -66,6 +66,14 @@
 #define XE_ARCH_PPC 1
 #endif
 
+#ifdef XE_ARCH_AMD64
+#define XE_HOST_ARCH_NAME "x64"
+#elif XE_ARCH_ARM64
+#define XE_HOST_ARCH_NAME "a64"
+#elif XE_ARCH_PPC
+#define XE_HOST_ARCH_NAME "ppc"
+#endif
+
 #if XE_PLATFORM_WIN32
 #define WIN32_LEAN_AND_MEAN
 #define NOMINMAX  // Don't want windows.h including min/max macros.