[dynarmic] minor optimization on codepage caching for non-fastmem

src/core/arm/dynarmic/arm_dynarmic_32.cpp

@@ -42,7 +42,7 @@ u64 DynarmicCallbacks32::MemoryRead64(u32 vaddr) {
 std::optional<u32> DynarmicCallbacks32::MemoryReadCode(u32 vaddr) {
     if (!m_memory.IsValidVirtualAddressRange(vaddr, sizeof(u32)))
         return std::nullopt;
-    auto const aligned_vaddr = vaddr & ~(Dynarmic::CODE_PAGE_SIZE - 1);
+    auto const aligned_vaddr = Common::AlignDown(vaddr, Dynarmic::CODE_PAGE_SIZE);
     if (last_code_addr != aligned_vaddr) {
         m_memory.ReadBlock(aligned_vaddr, &cached_code_page, sizeof(cached_code_page));
         last_code_addr = aligned_vaddr;

src/core/arm/dynarmic/arm_dynarmic_32.h

@@ -53,13 +53,13 @@ public:
     bool CheckMemoryAccess(u64 addr, u64 size, Kernel::DebugWatchpointType type);
     void ReturnException(u32 pc, Dynarmic::HaltReason hr);
     //
-    Dynarmic::CodePage cached_code_page;
     u64 last_code_addr = u64(-1);
     ArmDynarmic32& m_parent;
     Core::Memory::Memory& m_memory;
     Kernel::KProcess* m_process{};
     const bool m_debugger_enabled{};
     const bool m_check_memory_access{};
+    Dynarmic::CodePage cached_code_page;
 };
 
 class ArmDynarmic32 final : public ArmInterface {

src/core/arm/dynarmic/arm_dynarmic_64.cpp

@@ -45,7 +45,7 @@ Dynarmic::A64::Vector DynarmicCallbacks64::MemoryRead128(u64 vaddr) {
 std::optional<u32> DynarmicCallbacks64::MemoryReadCode(u64 vaddr) {
     if (!m_memory.IsValidVirtualAddressRange(vaddr, sizeof(u32)))
         return std::nullopt;
-    auto const aligned_vaddr = vaddr & ~(Dynarmic::CODE_PAGE_SIZE - 1);
+    auto const aligned_vaddr = Common::AlignDown(vaddr, Dynarmic::CODE_PAGE_SIZE);
     if (last_code_addr != aligned_vaddr) {
         m_memory.ReadBlock(aligned_vaddr, &cached_code_page, sizeof(cached_code_page));
         last_code_addr = aligned_vaddr;

src/core/arm/dynarmic/arm_dynarmic_64.h

@@ -64,7 +64,6 @@ public:
     bool CheckMemoryAccess(u64 addr, u64 size, Kernel::DebugWatchpointType type);
     void ReturnException(u64 pc, Dynarmic::HaltReason hr);
 
-    Dynarmic::CodePage cached_code_page;
     u64 last_code_addr = u64(-1);
     ArmDynarmic64& m_parent;
     Core::Memory::Memory& m_memory;
@@ -74,6 +73,7 @@ public:
     const bool m_debugger_enabled{};
     const bool m_check_memory_access{};
     static constexpr u64 MinimumRunCycles = 10000U;
+    Dynarmic::CodePage cached_code_page;
 };
 
 class ArmDynarmic64 final : public ArmInterface {

src/core/hle/service/jit/jit_context.cpp

@@ -60,9 +60,7 @@ public:
     {}
 
     std::optional<std::uint32_t> MemoryReadCode(VAddr vaddr) override {
-        // todo: does code page have to be 4kib?
-        //static_assert(Core::Memory::YUZU_PAGESIZE == Dynarmic::CODE_PAGE_SIZE);
-        auto const aligned_vaddr = vaddr & ~(Dynarmic::CODE_PAGE_SIZE - 1);
+        auto const aligned_vaddr = Common::AlignDown(vaddr, Dynarmic::CODE_PAGE_SIZE);
         if (last_code_addr != aligned_vaddr) {
             cached_code_page = ReadMemory<Dynarmic::CodePage>(aligned_vaddr);
             last_code_addr = aligned_vaddr;
@@ -167,8 +165,8 @@ private:
     std::vector<u8>& local_memory;
     IntervalSet& mapped_ranges;
     JITContextImpl& parent;
-    Dynarmic::CodePage cached_code_page;
     u64 last_code_addr = u64(-1);
+    Dynarmic::CodePage cached_code_page;
 };
 
 class JITContextImpl {

src/dynarmic/src/dynarmic/interface/code_page.h

@@ -8,10 +8,17 @@
 
 namespace Dynarmic {
 
-/// @brief Smallest valid page (may change for Apple?)
-constexpr inline uint64_t CODE_PAGE_SIZE = 4096;
+/// @brief Smallest valid page
+///
+// TODO: can we base this off the system page size without using the heap?
+#if defined(__APPLE__) && defined(__aarch64__)
+constexpr inline uint64_t CODE_PAGE_SIZE = 0x4000;
+#else
+constexpr inline uint64_t CODE_PAGE_SIZE = 0x1000;
+#endif
+
 struct CodePage {
-    uint32_t inst[CODE_PAGE_SIZE / sizeof(uint32_t)];
+    alignas(CODE_PAGE_SIZE) uint32_t inst[CODE_PAGE_SIZE / sizeof(uint32_t)];
 };
 
 }