use dmem for swap buffers, restore full jit sizes

src/common/virtual_buffer.cpp

@@ -16,6 +16,7 @@
 #include <fcntl.h>
 #include <boost/container/static_vector.hpp>
 #include <orbis/SystemService.h>
+#include <orbis/libkernel.h>
 typedef void (*SceKernelExceptionHandler)(int32_t, void*);
 extern "C" int32_t sceKernelInstallExceptionHandler(int32_t signum, SceKernelExceptionHandler handler);
 #endif
@@ -143,8 +144,13 @@ void* AllocateMemoryPages(std::size_t size) noexcept {
 #elif defined(__OPENORBIS__)
     void* addr;
     if (size <= 8192 * 4096) {
-        addr = malloc(size);
-        LOG_WARNING(HW_Memory, "Using DMem for {} bytes area @ {}", size, addr);
+        size_t align = 16384;
+        size_t len = (size + align - 1) / align * align;
+        off_t offset = 0;
+        ASSERT(sceKernelAllocateDirectMemory(0, ORBIS_KERNEL_MAIN_DMEM_SIZE, len, align, ORBIS_KERNEL_WB_ONION, &offset) == 0);
+        ASSERT(sceKernelMapDirectMemory(&addr, len, ORBIS_KERNEL_PROT_CPU_RW, 0, offset, len) == 0);
+        ASSERT(sceKernelMprotect(addr, len, VM_PROT_ALL) == 0);
+        LOG_WARNING(HW_Memory, "Using DMem for {} bytes area @ {}", len, addr);
         ASSERT(addr != nullptr);
     } else {
         addr = mmap(nullptr, size, PROT_READ | PROT_WRITE, MAP_VOID | MAP_PRIVATE, -1, 0);
@@ -166,7 +172,7 @@ void FreeMemoryPages(void* addr, [[maybe_unused]] std::size_t size) noexcept {
     VirtualFree(addr, 0, MEM_RELEASE);
 #elif defined(__OPENORBIS__)
     if (size <= 8192 * 4096) {
-        free(addr);
+        sceKernelCheckedReleaseDirectMemory(off_t(addr), size_t(size));
     } else {
         int rc = munmap(addr, size);
         ASSERT(rc == 0);

src/core/arm/dynarmic/arm_dynarmic_32.cpp

@@ -204,7 +204,7 @@ void ArmDynarmic32::MakeJit(Common::PageTable* page_table) {
 
     // Code cache size
 #if defined(__OPENORBIS__)
-    config.code_cache_size = std::uint32_t(32_MiB);
+    config.code_cache_size = std::uint32_t(128_MiB);
 #elif defined(ARCHITECTURE_arm64) || defined(__sun__) || defined(__NetBSD__) || defined(__DragonFly__) || defined(__OpenBSD__)
     config.code_cache_size = std::uint32_t(128_MiB);
 #else

src/core/arm/dynarmic/arm_dynarmic_64.cpp

@@ -255,7 +255,7 @@ void ArmDynarmic64::MakeJit(Common::PageTable* page_table, std::size_t address_s
 
     // Code cache size
 #if defined(__OPENORBIS__)
-    config.code_cache_size = std::uint32_t(32_MiB);
+    config.code_cache_size = std::uint32_t(128_MiB);
 #elif defined(ARCHITECTURE_arm64) || defined(__sun__) || defined(__NetBSD__) || defined(__DragonFly__) || defined(__OpenBSD__)
     config.code_cache_size = std::uint32_t(128_MiB);
 #else

src/dynarmic/src/dynarmic/backend/x64/block_of_code.cpp

@@ -58,13 +58,8 @@ const std::array<Xbyak::Reg64, ABI_PARAM_COUNT> BlockOfCode::ABI_PARAMS = {Block
 
 namespace {
 
-#ifdef __OPENORBIS__
-constexpr size_t CONSTANT_POOL_SIZE = 8 * 4096;
-constexpr size_t PRELUDE_COMMIT_SIZE = 8 * 4096;
-#else
 constexpr size_t CONSTANT_POOL_SIZE = 2 * 1024 * 1024;
 constexpr size_t PRELUDE_COMMIT_SIZE = 16 * 1024 * 1024;
-#endif
 
 class CustomXbyakAllocator : public Xbyak::Allocator {
 public:
@@ -247,9 +242,7 @@ bool IsUnderRosetta() {
 }  // anonymous namespace
 
 BlockOfCode::BlockOfCode(RunCodeCallbacks cb, JitStateInfo jsi, size_t total_code_size, std::function<void(BlockOfCode&)> rcp) noexcept
-#ifdef __OPENORBIS__
-    : Xbyak::CodeGenerator(total_code_size, Xbyak::AutoGrow, &s_allocator)
-#elif defined(DYNARMIC_ENABLE_NO_EXECUTE_SUPPORT)
+#if defined(DYNARMIC_ENABLE_NO_EXECUTE_SUPPORT)
     : Xbyak::CodeGenerator(total_code_size, Xbyak::DontSetProtectRWE, &s_allocator)
 #else
     : Xbyak::CodeGenerator(total_code_size, nullptr, &s_allocator)