summaryrefslogtreecommitdiffstats
path: root/src/core/arm/nce
diff options
context:
space:
mode:
Diffstat (limited to '')
-rw-r--r--src/core/arm/nce/arm_nce.cpp400
-rw-r--r--src/core/arm/nce/arm_nce.h108
-rw-r--r--src/core/arm/nce/arm_nce.s222
-rw-r--r--src/core/arm/nce/arm_nce_asm_definitions.h29
-rw-r--r--src/core/arm/nce/guest_context.h50
-rw-r--r--src/core/arm/nce/instructions.h147
-rw-r--r--src/core/arm/nce/patcher.cpp474
-rw-r--r--src/core/arm/nce/patcher.h98
8 files changed, 1528 insertions, 0 deletions
diff --git a/src/core/arm/nce/arm_nce.cpp b/src/core/arm/nce/arm_nce.cpp
new file mode 100644
index 000000000..f7bdafd39
--- /dev/null
+++ b/src/core/arm/nce/arm_nce.cpp
@@ -0,0 +1,400 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include <cinttypes>
+#include <memory>
+
+#include "common/signal_chain.h"
+#include "core/arm/nce/arm_nce.h"
+#include "core/arm/nce/patcher.h"
+#include "core/core.h"
+#include "core/memory.h"
+
+#include "core/hle/kernel/k_process.h"
+
+#include <signal.h>
+#include <sys/syscall.h>
+#include <unistd.h>
+
+namespace Core {
+
+namespace {
+
+struct sigaction g_orig_action;
+
+// Verify assembly offsets.
+using NativeExecutionParameters = Kernel::KThread::NativeExecutionParameters;
+static_assert(offsetof(NativeExecutionParameters, native_context) == TpidrEl0NativeContext);
+static_assert(offsetof(NativeExecutionParameters, lock) == TpidrEl0Lock);
+static_assert(offsetof(NativeExecutionParameters, magic) == TpidrEl0TlsMagic);
+
+fpsimd_context* GetFloatingPointState(mcontext_t& host_ctx) {
+ _aarch64_ctx* header = reinterpret_cast<_aarch64_ctx*>(&host_ctx.__reserved);
+ while (header->magic != FPSIMD_MAGIC) {
+ header = reinterpret_cast<_aarch64_ctx*>(reinterpret_cast<char*>(header) + header->size);
+ }
+ return reinterpret_cast<fpsimd_context*>(header);
+}
+
+} // namespace
+
+void* ARM_NCE::RestoreGuestContext(void* raw_context) {
+ // Retrieve the host context.
+ auto& host_ctx = static_cast<ucontext_t*>(raw_context)->uc_mcontext;
+
+ // Thread-local parameters will be located in x9.
+ auto* tpidr = reinterpret_cast<NativeExecutionParameters*>(host_ctx.regs[9]);
+ auto* guest_ctx = static_cast<GuestContext*>(tpidr->native_context);
+
+ // Retrieve the host floating point state.
+ auto* fpctx = GetFloatingPointState(host_ctx);
+
+ // Save host callee-saved registers.
+ std::memcpy(guest_ctx->host_ctx.host_saved_vregs.data(), &fpctx->vregs[8],
+ sizeof(guest_ctx->host_ctx.host_saved_vregs));
+ std::memcpy(guest_ctx->host_ctx.host_saved_regs.data(), &host_ctx.regs[19],
+ sizeof(guest_ctx->host_ctx.host_saved_regs));
+
+ // Save stack pointer.
+ guest_ctx->host_ctx.host_sp = host_ctx.sp;
+
+ // Restore all guest state except tpidr_el0.
+ host_ctx.sp = guest_ctx->sp;
+ host_ctx.pc = guest_ctx->pc;
+ host_ctx.pstate = guest_ctx->pstate;
+ fpctx->fpcr = guest_ctx->fpcr;
+ fpctx->fpsr = guest_ctx->fpsr;
+ std::memcpy(host_ctx.regs, guest_ctx->cpu_registers.data(), sizeof(host_ctx.regs));
+ std::memcpy(fpctx->vregs, guest_ctx->vector_registers.data(), sizeof(fpctx->vregs));
+
+ // Return the new thread-local storage pointer.
+ return tpidr;
+}
+
+void ARM_NCE::SaveGuestContext(GuestContext* guest_ctx, void* raw_context) {
+ // Retrieve the host context.
+ auto& host_ctx = static_cast<ucontext_t*>(raw_context)->uc_mcontext;
+
+ // Retrieve the host floating point state.
+ auto* fpctx = GetFloatingPointState(host_ctx);
+
+ // Save all guest registers except tpidr_el0.
+ std::memcpy(guest_ctx->cpu_registers.data(), host_ctx.regs, sizeof(host_ctx.regs));
+ std::memcpy(guest_ctx->vector_registers.data(), fpctx->vregs, sizeof(fpctx->vregs));
+ guest_ctx->fpsr = fpctx->fpsr;
+ guest_ctx->fpcr = fpctx->fpcr;
+ guest_ctx->pstate = static_cast<u32>(host_ctx.pstate);
+ guest_ctx->pc = host_ctx.pc;
+ guest_ctx->sp = host_ctx.sp;
+
+ // Restore stack pointer.
+ host_ctx.sp = guest_ctx->host_ctx.host_sp;
+
+ // Restore host callee-saved registers.
+ std::memcpy(&host_ctx.regs[19], guest_ctx->host_ctx.host_saved_regs.data(),
+ sizeof(guest_ctx->host_ctx.host_saved_regs));
+ std::memcpy(&fpctx->vregs[8], guest_ctx->host_ctx.host_saved_vregs.data(),
+ sizeof(guest_ctx->host_ctx.host_saved_vregs));
+
+ // Return from the call on exit by setting pc to x30.
+ host_ctx.pc = guest_ctx->host_ctx.host_saved_regs[11];
+
+ // Clear esr_el1 and return it.
+ host_ctx.regs[0] = guest_ctx->esr_el1.exchange(0);
+}
+
+bool ARM_NCE::HandleGuestFault(GuestContext* guest_ctx, void* raw_info, void* raw_context) {
+ auto& host_ctx = static_cast<ucontext_t*>(raw_context)->uc_mcontext;
+ auto* info = static_cast<siginfo_t*>(raw_info);
+
+ // Try to handle an invalid access.
+ // TODO: handle accesses which split a page?
+ const Common::ProcessAddress addr =
+ (reinterpret_cast<u64>(info->si_addr) & ~Memory::YUZU_PAGEMASK);
+ if (guest_ctx->system->ApplicationMemory().InvalidateNCE(addr, Memory::YUZU_PAGESIZE)) {
+ // We handled the access successfully and are returning to guest code.
+ return true;
+ }
+
+ // We can't handle the access, so determine why we crashed.
+ const bool is_prefetch_abort = host_ctx.pc == reinterpret_cast<u64>(info->si_addr);
+
+ // For data aborts, skip the instruction and return to guest code.
+ // This will allow games to continue in many scenarios where they would otherwise crash.
+ if (!is_prefetch_abort) {
+ host_ctx.pc += 4;
+ return true;
+ }
+
+ // This is a prefetch abort.
+ guest_ctx->esr_el1.fetch_or(static_cast<u64>(HaltReason::PrefetchAbort));
+
+ // Forcibly mark the context as locked. We are still running.
+ // We may race with SignalInterrupt here:
+ // - If we lose the race, then SignalInterrupt will send us a signal we are masking,
+ // and it will do nothing when it is unmasked, as we have already left guest code.
+ // - If we win the race, then SignalInterrupt will wait for us to unlock first.
+ auto& thread_params = guest_ctx->parent->running_thread->GetNativeExecutionParameters();
+ thread_params.lock.store(SpinLockLocked);
+
+ // Return to host.
+ SaveGuestContext(guest_ctx, raw_context);
+ return false;
+}
+
+void ARM_NCE::HandleHostFault(int sig, void* raw_info, void* raw_context) {
+ return g_orig_action.sa_sigaction(sig, static_cast<siginfo_t*>(raw_info), raw_context);
+}
+
+HaltReason ARM_NCE::RunJit() {
+ // Get the thread parameters.
+ // TODO: pass the current thread down from ::Run
+ auto* thread = Kernel::GetCurrentThreadPointer(system.Kernel());
+ auto* thread_params = &thread->GetNativeExecutionParameters();
+
+ {
+ // Lock our core context.
+ std::scoped_lock lk{lock};
+
+ // We should not be running.
+ ASSERT(running_thread == nullptr);
+
+ // Check if we need to run. If we have already been halted, we are done.
+ u64 halt = guest_ctx.esr_el1.exchange(0);
+ if (halt != 0) {
+ return static_cast<HaltReason>(halt);
+ }
+
+ // Mark that we are running.
+ running_thread = thread;
+
+ // Acquire the lock on the thread parameters.
+ // This allows us to force synchronization with SignalInterrupt.
+ LockThreadParameters(thread_params);
+ }
+
+ // Assign current members.
+ guest_ctx.parent = this;
+ thread_params->native_context = &guest_ctx;
+ thread_params->tpidr_el0 = guest_ctx.tpidr_el0;
+ thread_params->tpidrro_el0 = guest_ctx.tpidrro_el0;
+ thread_params->is_running = true;
+
+ HaltReason halt{};
+
+ // TODO: finding and creating the post handler needs to be locked
+ // to deal with dynamic loading of NROs.
+ const auto& post_handlers = system.ApplicationProcess()->GetPostHandlers();
+ if (auto it = post_handlers.find(guest_ctx.pc); it != post_handlers.end()) {
+ halt = ReturnToRunCodeByTrampoline(thread_params, &guest_ctx, it->second);
+ } else {
+ halt = ReturnToRunCodeByExceptionLevelChange(thread_id, thread_params);
+ }
+
+ // Unload members.
+ // The thread does not change, so we can persist the old reference.
+ guest_ctx.tpidr_el0 = thread_params->tpidr_el0;
+ thread_params->native_context = nullptr;
+ thread_params->is_running = false;
+
+ // Unlock the thread parameters.
+ UnlockThreadParameters(thread_params);
+
+ {
+ // Lock the core context.
+ std::scoped_lock lk{lock};
+
+ // On exit, we no longer have an active thread.
+ running_thread = nullptr;
+ }
+
+ // Return the halt reason.
+ return halt;
+}
+
+HaltReason ARM_NCE::StepJit() {
+ return HaltReason::StepThread;
+}
+
+u32 ARM_NCE::GetSvcNumber() const {
+ return guest_ctx.svc_swi;
+}
+
+ARM_NCE::ARM_NCE(System& system_, bool uses_wall_clock_, std::size_t core_index_)
+ : ARM_Interface{system_, uses_wall_clock_}, core_index{core_index_} {
+ guest_ctx.system = &system_;
+}
+
+ARM_NCE::~ARM_NCE() = default;
+
+void ARM_NCE::Initialize() {
+ thread_id = gettid();
+
+ // Setup our signals
+ static std::once_flag flag;
+ std::call_once(flag, [] {
+ using HandlerType = decltype(sigaction::sa_sigaction);
+
+ sigset_t signal_mask;
+ sigemptyset(&signal_mask);
+ sigaddset(&signal_mask, ReturnToRunCodeByExceptionLevelChangeSignal);
+ sigaddset(&signal_mask, BreakFromRunCodeSignal);
+ sigaddset(&signal_mask, GuestFaultSignal);
+
+ struct sigaction return_to_run_code_action {};
+ return_to_run_code_action.sa_flags = SA_SIGINFO | SA_ONSTACK;
+ return_to_run_code_action.sa_sigaction = reinterpret_cast<HandlerType>(
+ &ARM_NCE::ReturnToRunCodeByExceptionLevelChangeSignalHandler);
+ return_to_run_code_action.sa_mask = signal_mask;
+ Common::SigAction(ReturnToRunCodeByExceptionLevelChangeSignal, &return_to_run_code_action,
+ nullptr);
+
+ struct sigaction break_from_run_code_action {};
+ break_from_run_code_action.sa_flags = SA_SIGINFO | SA_ONSTACK;
+ break_from_run_code_action.sa_sigaction =
+ reinterpret_cast<HandlerType>(&ARM_NCE::BreakFromRunCodeSignalHandler);
+ break_from_run_code_action.sa_mask = signal_mask;
+ Common::SigAction(BreakFromRunCodeSignal, &break_from_run_code_action, nullptr);
+
+ struct sigaction fault_action {};
+ fault_action.sa_flags = SA_SIGINFO | SA_ONSTACK | SA_RESTART;
+ fault_action.sa_sigaction =
+ reinterpret_cast<HandlerType>(&ARM_NCE::GuestFaultSignalHandler);
+ fault_action.sa_mask = signal_mask;
+ Common::SigAction(GuestFaultSignal, &fault_action, &g_orig_action);
+
+ // Simplify call for g_orig_action.
+ // These fields occupy the same space in memory, so this should be a no-op in practice.
+ if (!(g_orig_action.sa_flags & SA_SIGINFO)) {
+ g_orig_action.sa_sigaction =
+ reinterpret_cast<decltype(g_orig_action.sa_sigaction)>(g_orig_action.sa_handler);
+ }
+ });
+}
+
+void ARM_NCE::SetPC(u64 pc) {
+ guest_ctx.pc = pc;
+}
+
+u64 ARM_NCE::GetPC() const {
+ return guest_ctx.pc;
+}
+
+u64 ARM_NCE::GetSP() const {
+ return guest_ctx.sp;
+}
+
+u64 ARM_NCE::GetReg(int index) const {
+ return guest_ctx.cpu_registers[index];
+}
+
+void ARM_NCE::SetReg(int index, u64 value) {
+ guest_ctx.cpu_registers[index] = value;
+}
+
+u128 ARM_NCE::GetVectorReg(int index) const {
+ return guest_ctx.vector_registers[index];
+}
+
+void ARM_NCE::SetVectorReg(int index, u128 value) {
+ guest_ctx.vector_registers[index] = value;
+}
+
+u32 ARM_NCE::GetPSTATE() const {
+ return guest_ctx.pstate;
+}
+
+void ARM_NCE::SetPSTATE(u32 pstate) {
+ guest_ctx.pstate = pstate;
+}
+
+u64 ARM_NCE::GetTlsAddress() const {
+ return guest_ctx.tpidrro_el0;
+}
+
+void ARM_NCE::SetTlsAddress(u64 address) {
+ guest_ctx.tpidrro_el0 = address;
+}
+
+u64 ARM_NCE::GetTPIDR_EL0() const {
+ return guest_ctx.tpidr_el0;
+}
+
+void ARM_NCE::SetTPIDR_EL0(u64 value) {
+ guest_ctx.tpidr_el0 = value;
+}
+
+void ARM_NCE::SaveContext(ThreadContext64& ctx) const {
+ ctx.cpu_registers = guest_ctx.cpu_registers;
+ ctx.sp = guest_ctx.sp;
+ ctx.pc = guest_ctx.pc;
+ ctx.pstate = guest_ctx.pstate;
+ ctx.vector_registers = guest_ctx.vector_registers;
+ ctx.fpcr = guest_ctx.fpcr;
+ ctx.fpsr = guest_ctx.fpsr;
+ ctx.tpidr = guest_ctx.tpidr_el0;
+}
+
+void ARM_NCE::LoadContext(const ThreadContext64& ctx) {
+ guest_ctx.cpu_registers = ctx.cpu_registers;
+ guest_ctx.sp = ctx.sp;
+ guest_ctx.pc = ctx.pc;
+ guest_ctx.pstate = ctx.pstate;
+ guest_ctx.vector_registers = ctx.vector_registers;
+ guest_ctx.fpcr = ctx.fpcr;
+ guest_ctx.fpsr = ctx.fpsr;
+ guest_ctx.tpidr_el0 = ctx.tpidr;
+}
+
+void ARM_NCE::SignalInterrupt() {
+ // Lock core context.
+ std::scoped_lock lk{lock};
+
+ // Add break loop condition.
+ guest_ctx.esr_el1.fetch_or(static_cast<u64>(HaltReason::BreakLoop));
+
+ // If there is no thread running, we are done.
+ if (running_thread == nullptr) {
+ return;
+ }
+
+ // Lock the thread context.
+ auto* params = &running_thread->GetNativeExecutionParameters();
+ LockThreadParameters(params);
+
+ if (params->is_running) {
+ // We should signal to the running thread.
+ // The running thread will unlock the thread context.
+ syscall(SYS_tkill, thread_id, BreakFromRunCodeSignal);
+ } else {
+ // If the thread is no longer running, we have nothing to do.
+ UnlockThreadParameters(params);
+ }
+}
+
+void ARM_NCE::ClearInterrupt() {
+ guest_ctx.esr_el1 = {};
+}
+
+void ARM_NCE::ClearInstructionCache() {
+ // TODO: This is not possible to implement correctly on Linux because
+ // we do not have any access to ic iallu.
+
+ // Require accesses to complete.
+ std::atomic_thread_fence(std::memory_order_seq_cst);
+}
+
+void ARM_NCE::InvalidateCacheRange(u64 addr, std::size_t size) {
+ this->ClearInstructionCache();
+}
+
+void ARM_NCE::ClearExclusiveState() {
+ // No-op.
+}
+
+void ARM_NCE::PageTableChanged(Common::PageTable& page_table,
+ std::size_t new_address_space_size_in_bits) {
+ // No-op. Page table is never used.
+}
+
+} // namespace Core
diff --git a/src/core/arm/nce/arm_nce.h b/src/core/arm/nce/arm_nce.h
new file mode 100644
index 000000000..5fbd6dbf3
--- /dev/null
+++ b/src/core/arm/nce/arm_nce.h
@@ -0,0 +1,108 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <atomic>
+#include <memory>
+#include <span>
+#include <unordered_map>
+#include <vector>
+
+#include "core/arm/arm_interface.h"
+#include "core/arm/nce/guest_context.h"
+
+namespace Core::Memory {
+class Memory;
+}
+
+namespace Core {
+
+class System;
+
+class ARM_NCE final : public ARM_Interface {
+public:
+ ARM_NCE(System& system_, bool uses_wall_clock_, std::size_t core_index_);
+
+ ~ARM_NCE() override;
+
+ void Initialize() override;
+ void SetPC(u64 pc) override;
+ u64 GetPC() const override;
+ u64 GetSP() const override;
+ u64 GetReg(int index) const override;
+ void SetReg(int index, u64 value) override;
+ u128 GetVectorReg(int index) const override;
+ void SetVectorReg(int index, u128 value) override;
+
+ u32 GetPSTATE() const override;
+ void SetPSTATE(u32 pstate) override;
+ u64 GetTlsAddress() const override;
+ void SetTlsAddress(u64 address) override;
+ void SetTPIDR_EL0(u64 value) override;
+ u64 GetTPIDR_EL0() const override;
+
+ Architecture GetArchitecture() const override {
+ return Architecture::Aarch64;
+ }
+
+ void SaveContext(ThreadContext32& ctx) const override {}
+ void SaveContext(ThreadContext64& ctx) const override;
+ void LoadContext(const ThreadContext32& ctx) override {}
+ void LoadContext(const ThreadContext64& ctx) override;
+
+ void SignalInterrupt() override;
+ void ClearInterrupt() override;
+ void ClearExclusiveState() override;
+ void ClearInstructionCache() override;
+ void InvalidateCacheRange(u64 addr, std::size_t size) override;
+ void PageTableChanged(Common::PageTable& new_page_table,
+ std::size_t new_address_space_size_in_bits) override;
+
+protected:
+ HaltReason RunJit() override;
+ HaltReason StepJit() override;
+
+ u32 GetSvcNumber() const override;
+
+ const Kernel::DebugWatchpoint* HaltedWatchpoint() const override {
+ return nullptr;
+ }
+
+ void RewindBreakpointInstruction() override {}
+
+private:
+ // Assembly definitions.
+ static HaltReason ReturnToRunCodeByTrampoline(void* tpidr, GuestContext* ctx,
+ u64 trampoline_addr);
+ static HaltReason ReturnToRunCodeByExceptionLevelChange(int tid, void* tpidr);
+
+ static void ReturnToRunCodeByExceptionLevelChangeSignalHandler(int sig, void* info,
+ void* raw_context);
+ static void BreakFromRunCodeSignalHandler(int sig, void* info, void* raw_context);
+ static void GuestFaultSignalHandler(int sig, void* info, void* raw_context);
+
+ static void LockThreadParameters(void* tpidr);
+ static void UnlockThreadParameters(void* tpidr);
+
+private:
+ // C++ implementation functions for assembly definitions.
+ static void* RestoreGuestContext(void* raw_context);
+ static void SaveGuestContext(GuestContext* ctx, void* raw_context);
+ static bool HandleGuestFault(GuestContext* ctx, void* info, void* raw_context);
+ static void HandleHostFault(int sig, void* info, void* raw_context);
+
+public:
+ // Members set on initialization.
+ std::size_t core_index{};
+ pid_t thread_id{-1};
+
+ // Core context.
+ GuestContext guest_ctx;
+
+ // Thread and invalidation info.
+ std::mutex lock;
+ Kernel::KThread* running_thread{};
+};
+
+} // namespace Core
diff --git a/src/core/arm/nce/arm_nce.s b/src/core/arm/nce/arm_nce.s
new file mode 100644
index 000000000..b98e09f31
--- /dev/null
+++ b/src/core/arm/nce/arm_nce.s
@@ -0,0 +1,222 @@
+/* SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project */
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+
+#include "core/arm/nce/arm_nce_asm_definitions.h"
+
+#define LOAD_IMMEDIATE_32(reg, val) \
+ mov reg, #(((val) >> 0x00) & 0xFFFF); \
+ movk reg, #(((val) >> 0x10) & 0xFFFF), lsl #16
+
+
+/* static HaltReason Core::ARM_NCE::ReturnToRunCodeByTrampoline(void* tpidr, Core::GuestContext* ctx, u64 trampoline_addr) */
+.section .text._ZN4Core7ARM_NCE27ReturnToRunCodeByTrampolineEPvPNS_12GuestContextEm, "ax", %progbits
+.global _ZN4Core7ARM_NCE27ReturnToRunCodeByTrampolineEPvPNS_12GuestContextEm
+.type _ZN4Core7ARM_NCE27ReturnToRunCodeByTrampolineEPvPNS_12GuestContextEm, %function
+_ZN4Core7ARM_NCE27ReturnToRunCodeByTrampolineEPvPNS_12GuestContextEm:
+ /* Back up host sp to x3. */
+ /* Back up host tpidr_el0 to x4. */
+ mov x3, sp
+ mrs x4, tpidr_el0
+
+ /* Load guest sp. x5 is used as a scratch register. */
+ ldr x5, [x1, #(GuestContextSp)]
+ mov sp, x5
+
+ /* Offset GuestContext pointer to the host member. */
+ add x5, x1, #(GuestContextHostContext)
+
+ /* Save original host sp and tpidr_el0 (x3, x4) to host context. */
+ stp x3, x4, [x5, #(HostContextSpTpidrEl0)]
+
+ /* Save all callee-saved host GPRs. */
+ stp x19, x20, [x5, #(HostContextRegs+0x0)]
+ stp x21, x22, [x5, #(HostContextRegs+0x10)]
+ stp x23, x24, [x5, #(HostContextRegs+0x20)]
+ stp x25, x26, [x5, #(HostContextRegs+0x30)]
+ stp x27, x28, [x5, #(HostContextRegs+0x40)]
+ stp x29, x30, [x5, #(HostContextRegs+0x50)]
+
+ /* Save all callee-saved host FPRs. */
+ stp q8, q9, [x5, #(HostContextVregs+0x0)]
+ stp q10, q11, [x5, #(HostContextVregs+0x20)]
+ stp q12, q13, [x5, #(HostContextVregs+0x40)]
+ stp q14, q15, [x5, #(HostContextVregs+0x60)]
+
+ /* Load guest tpidr_el0 from argument. */
+ msr tpidr_el0, x0
+
+ /* Tail call the trampoline to restore guest state. */
+ br x2
+
+
+/* static HaltReason Core::ARM_NCE::ReturnToRunCodeByExceptionLevelChange(int tid, void* tpidr) */
+.section .text._ZN4Core7ARM_NCE37ReturnToRunCodeByExceptionLevelChangeEiPv, "ax", %progbits
+.global _ZN4Core7ARM_NCE37ReturnToRunCodeByExceptionLevelChangeEiPv
+.type _ZN4Core7ARM_NCE37ReturnToRunCodeByExceptionLevelChangeEiPv, %function
+_ZN4Core7ARM_NCE37ReturnToRunCodeByExceptionLevelChangeEiPv:
+ /* This jumps to the signal handler, which will restore the entire context. */
+ /* On entry, x0 = thread id, which is already in the right place. */
+
+ /* Move tpidr to x9 so it is not trampled. */
+ mov x9, x1
+
+ /* Set up arguments. */
+ mov x8, #(__NR_tkill)
+ mov x1, #(ReturnToRunCodeByExceptionLevelChangeSignal)
+
+ /* Tail call the signal handler. */
+ svc #0
+
+ /* Block execution from flowing here. */
+ brk #1000
+
+
+/* static void Core::ARM_NCE::ReturnToRunCodeByExceptionLevelChangeSignalHandler(int sig, void* info, void* raw_context) */
+.section .text._ZN4Core7ARM_NCE50ReturnToRunCodeByExceptionLevelChangeSignalHandlerEiPvS1_, "ax", %progbits
+.global _ZN4Core7ARM_NCE50ReturnToRunCodeByExceptionLevelChangeSignalHandlerEiPvS1_
+.type _ZN4Core7ARM_NCE50ReturnToRunCodeByExceptionLevelChangeSignalHandlerEiPvS1_, %function
+_ZN4Core7ARM_NCE50ReturnToRunCodeByExceptionLevelChangeSignalHandlerEiPvS1_:
+ stp x29, x30, [sp, #-0x10]!
+ mov x29, sp
+
+ /* Call the context restorer with the raw context. */
+ mov x0, x2
+ bl _ZN4Core7ARM_NCE19RestoreGuestContextEPv
+
+ /* Save the old value of tpidr_el0. */
+ mrs x8, tpidr_el0
+ ldr x9, [x0, #(TpidrEl0NativeContext)]
+ str x8, [x9, #(GuestContextHostContext + HostContextTpidrEl0)]
+
+ /* Set our new tpidr_el0. */
+ msr tpidr_el0, x0
+
+ /* Unlock the context. */
+ bl _ZN4Core7ARM_NCE22UnlockThreadParametersEPv
+
+ /* Returning from here will enter the guest. */
+ ldp x29, x30, [sp], #0x10
+ ret
+
+
+/* static void Core::ARM_NCE::BreakFromRunCodeSignalHandler(int sig, void* info, void* raw_context) */
+.section .text._ZN4Core7ARM_NCE29BreakFromRunCodeSignalHandlerEiPvS1_, "ax", %progbits
+.global _ZN4Core7ARM_NCE29BreakFromRunCodeSignalHandlerEiPvS1_
+.type _ZN4Core7ARM_NCE29BreakFromRunCodeSignalHandlerEiPvS1_, %function
+_ZN4Core7ARM_NCE29BreakFromRunCodeSignalHandlerEiPvS1_:
+ /* Check to see if we have the correct TLS magic. */
+ mrs x8, tpidr_el0
+ ldr w9, [x8, #(TpidrEl0TlsMagic)]
+
+ LOAD_IMMEDIATE_32(w10, TlsMagic)
+
+ cmp w9, w10
+ b.ne 1f
+
+ /* Correct TLS magic, so this is a guest interrupt. */
+ /* Restore host tpidr_el0. */
+ ldr x0, [x8, #(TpidrEl0NativeContext)]
+ ldr x3, [x0, #(GuestContextHostContext + HostContextTpidrEl0)]
+ msr tpidr_el0, x3
+
+ /* Tail call the restorer. */
+ mov x1, x2
+ b _ZN4Core7ARM_NCE16SaveGuestContextEPNS_12GuestContextEPv
+
+ /* Returning from here will enter host code. */
+
+1:
+ /* Incorrect TLS magic, so this is a spurious signal. */
+ ret
+
+
+/* static void Core::ARM_NCE::GuestFaultSignalHandler(int sig, void* info, void* raw_context) */
+.section .text._ZN4Core7ARM_NCE23GuestFaultSignalHandlerEiPvS1_, "ax", %progbits
+.global _ZN4Core7ARM_NCE23GuestFaultSignalHandlerEiPvS1_
+.type _ZN4Core7ARM_NCE23GuestFaultSignalHandlerEiPvS1_, %function
+_ZN4Core7ARM_NCE23GuestFaultSignalHandlerEiPvS1_:
+ /* Check to see if we have the correct TLS magic. */
+ mrs x8, tpidr_el0
+ ldr w9, [x8, #(TpidrEl0TlsMagic)]
+
+ LOAD_IMMEDIATE_32(w10, TlsMagic)
+
+ cmp w9, w10
+ b.eq 1f
+
+ /* Incorrect TLS magic, so this is a host fault. */
+ /* Tail call the handler. */
+ b _ZN4Core7ARM_NCE15HandleHostFaultEiPvS1_
+
+1:
+ /* Correct TLS magic, so this is a guest fault. */
+ stp x29, x30, [sp, #-0x20]!
+ str x19, [sp, #0x10]
+ mov x29, sp
+
+ /* Save the old tpidr_el0. */
+ mov x19, x8
+
+ /* Restore host tpidr_el0. */
+ ldr x0, [x8, #(TpidrEl0NativeContext)]
+ ldr x3, [x0, #(GuestContextHostContext + HostContextTpidrEl0)]
+ msr tpidr_el0, x3
+
+ /* Call the handler. */
+ bl _ZN4Core7ARM_NCE16HandleGuestFaultEPNS_12GuestContextEPvS3_
+
+ /* If the handler returned false, we want to preserve the host tpidr_el0. */
+ cbz x0, 2f
+
+ /* Otherwise, restore guest tpidr_el0. */
+ msr tpidr_el0, x19
+
+2:
+ ldr x19, [sp, #0x10]
+ ldp x29, x30, [sp], #0x20
+ ret
+
+
+/* static void Core::ARM_NCE::LockThreadParameters(void* tpidr) */
+.section .text._ZN4Core7ARM_NCE20LockThreadParametersEPv, "ax", %progbits
+.global _ZN4Core7ARM_NCE20LockThreadParametersEPv
+.type _ZN4Core7ARM_NCE20LockThreadParametersEPv, %function
+_ZN4Core7ARM_NCE20LockThreadParametersEPv:
+ /* Offset to lock member. */
+ add x0, x0, #(TpidrEl0Lock)
+
+1:
+ /* Clear the monitor. */
+ clrex
+
+2:
+ /* Load-linked with acquire ordering. */
+ ldaxr w1, [x0]
+
+ /* If the value was SpinLockLocked, clear monitor and retry. */
+ cbz w1, 1b
+
+ /* Store-conditional SpinLockLocked with relaxed ordering. */
+ stxr w1, wzr, [x0]
+
+ /* If we failed to store, retry. */
+ cbnz w1, 2b
+
+ ret
+
+
+/* static void Core::ARM_NCE::UnlockThreadParameters(void* tpidr) */
+.section .text._ZN4Core7ARM_NCE22UnlockThreadParametersEPv, "ax", %progbits
+.global _ZN4Core7ARM_NCE22UnlockThreadParametersEPv
+.type _ZN4Core7ARM_NCE22UnlockThreadParametersEPv, %function
+_ZN4Core7ARM_NCE22UnlockThreadParametersEPv:
+ /* Offset to lock member. */
+ add x0, x0, #(TpidrEl0Lock)
+
+ /* Load SpinLockUnlocked. */
+ mov w1, #(SpinLockUnlocked)
+
+ /* Store value with release ordering. */
+ stlr w1, [x0]
+
+ ret
diff --git a/src/core/arm/nce/arm_nce_asm_definitions.h b/src/core/arm/nce/arm_nce_asm_definitions.h
new file mode 100644
index 000000000..8a9b285b5
--- /dev/null
+++ b/src/core/arm/nce/arm_nce_asm_definitions.h
@@ -0,0 +1,29 @@
+/* SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project */
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+
+#pragma once
+
+#define __ASSEMBLY__
+
+#include <asm-generic/signal.h>
+#include <asm-generic/unistd.h>
+
+#define ReturnToRunCodeByExceptionLevelChangeSignal SIGUSR2
+#define BreakFromRunCodeSignal SIGURG
+#define GuestFaultSignal SIGSEGV
+
+#define GuestContextSp 0xF8
+#define GuestContextHostContext 0x320
+
+#define HostContextSpTpidrEl0 0xE0
+#define HostContextTpidrEl0 0xE8
+#define HostContextRegs 0x0
+#define HostContextVregs 0x60
+
+#define TpidrEl0NativeContext 0x10
+#define TpidrEl0Lock 0x18
+#define TpidrEl0TlsMagic 0x20
+#define TlsMagic 0x555a5559
+
+#define SpinLockLocked 0
+#define SpinLockUnlocked 1
diff --git a/src/core/arm/nce/guest_context.h b/src/core/arm/nce/guest_context.h
new file mode 100644
index 000000000..0767a0337
--- /dev/null
+++ b/src/core/arm/nce/guest_context.h
@@ -0,0 +1,50 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include "common/common_funcs.h"
+#include "common/common_types.h"
+#include "core/arm/arm_interface.h"
+#include "core/arm/nce/arm_nce_asm_definitions.h"
+
+namespace Core {
+
+class ARM_NCE;
+class System;
+
+struct HostContext {
+ alignas(16) std::array<u64, 12> host_saved_regs{};
+ alignas(16) std::array<u128, 8> host_saved_vregs{};
+ u64 host_sp{};
+ void* host_tpidr_el0{};
+};
+
+struct GuestContext {
+ std::array<u64, 31> cpu_registers{};
+ u64 sp{};
+ u64 pc{};
+ u32 fpcr{};
+ u32 fpsr{};
+ std::array<u128, 32> vector_registers{};
+ u32 pstate{};
+ alignas(16) HostContext host_ctx{};
+ u64 tpidrro_el0{};
+ u64 tpidr_el0{};
+ std::atomic<u64> esr_el1{};
+ u32 nzcv{};
+ u32 svc_swi{};
+ System* system{};
+ ARM_NCE* parent{};
+};
+
+// Verify assembly offsets.
+static_assert(offsetof(GuestContext, sp) == GuestContextSp);
+static_assert(offsetof(GuestContext, host_ctx) == GuestContextHostContext);
+static_assert(offsetof(HostContext, host_sp) == HostContextSpTpidrEl0);
+static_assert(offsetof(HostContext, host_tpidr_el0) - 8 == HostContextSpTpidrEl0);
+static_assert(offsetof(HostContext, host_tpidr_el0) == HostContextTpidrEl0);
+static_assert(offsetof(HostContext, host_saved_regs) == HostContextRegs);
+static_assert(offsetof(HostContext, host_saved_vregs) == HostContextVregs);
+
+} // namespace Core
diff --git a/src/core/arm/nce/instructions.h b/src/core/arm/nce/instructions.h
new file mode 100644
index 000000000..5b56ff857
--- /dev/null
+++ b/src/core/arm/nce/instructions.h
@@ -0,0 +1,147 @@
+// SPDX-FileCopyrightText: Copyright © 2020 Skyline Team and Contributors
+// SPDX-License-Identifier: MPL-2.0
+
+#include "common/bit_field.h"
+#include "common/common_types.h"
+
+namespace Core::NCE {
+
+enum SystemRegister : u32 {
+ TpidrEl0 = 0x5E82,
+ TpidrroEl0 = 0x5E83,
+ CntfrqEl0 = 0x5F00,
+ CntpctEl0 = 0x5F01,
+};
+
+// https://developer.arm.com/documentation/ddi0596/2021-12/Base-Instructions/SVC--Supervisor-Call-
+union SVC {
+ constexpr explicit SVC(u32 raw_) : raw{raw_} {}
+
+ constexpr bool Verify() {
+ return (this->GetSig0() == 0x1 && this->GetSig1() == 0x6A0);
+ }
+
+ constexpr u32 GetSig0() {
+ return decltype(sig0)::ExtractValue(raw);
+ }
+
+ constexpr u32 GetValue() {
+ return decltype(value)::ExtractValue(raw);
+ }
+
+ constexpr u32 GetSig1() {
+ return decltype(sig1)::ExtractValue(raw);
+ }
+
+ u32 raw;
+
+private:
+ BitField<0, 5, u32> sig0; // 0x1
+ BitField<5, 16, u32> value; // 16-bit immediate
+ BitField<21, 11, u32> sig1; // 0x6A0
+};
+static_assert(sizeof(SVC) == sizeof(u32));
+static_assert(SVC(0xD40000C1).Verify());
+static_assert(SVC(0xD40000C1).GetValue() == 0x6);
+
+// https://developer.arm.com/documentation/ddi0596/2021-12/Base-Instructions/MRS--Move-System-Register-
+union MRS {
+ constexpr explicit MRS(u32 raw_) : raw{raw_} {}
+
+ constexpr bool Verify() {
+ return (this->GetSig() == 0xD53);
+ }
+
+ constexpr u32 GetRt() {
+ return decltype(rt)::ExtractValue(raw);
+ }
+
+ constexpr u32 GetSystemReg() {
+ return decltype(system_reg)::ExtractValue(raw);
+ }
+
+ constexpr u32 GetSig() {
+ return decltype(sig)::ExtractValue(raw);
+ }
+
+ u32 raw;
+
+private:
+ BitField<0, 5, u32> rt; // destination register
+ BitField<5, 15, u32> system_reg; // source system register
+ BitField<20, 12, u32> sig; // 0xD53
+};
+static_assert(sizeof(MRS) == sizeof(u32));
+static_assert(MRS(0xD53BE020).Verify());
+static_assert(MRS(0xD53BE020).GetSystemReg() == CntpctEl0);
+static_assert(MRS(0xD53BE020).GetRt() == 0x0);
+
+// https://developer.arm.com/documentation/ddi0596/2021-12/Base-Instructions/MSR--register---Move-general-purpose-register-to-System-Register-
+union MSR {
+ constexpr explicit MSR(u32 raw_) : raw{raw_} {}
+
+ constexpr bool Verify() {
+ return this->GetSig() == 0xD51;
+ }
+
+ constexpr u32 GetRt() {
+ return decltype(rt)::ExtractValue(raw);
+ }
+
+ constexpr u32 GetSystemReg() {
+ return decltype(system_reg)::ExtractValue(raw);
+ }
+
+ constexpr u32 GetSig() {
+ return decltype(sig)::ExtractValue(raw);
+ }
+
+ u32 raw;
+
+private:
+ BitField<0, 5, u32> rt; // source register
+ BitField<5, 15, u32> system_reg; // destination system register
+ BitField<20, 12, u32> sig; // 0xD51
+};
+static_assert(sizeof(MSR) == sizeof(u32));
+static_assert(MSR(0xD51BD040).Verify());
+static_assert(MSR(0xD51BD040).GetSystemReg() == TpidrEl0);
+static_assert(MSR(0xD51BD040).GetRt() == 0x0);
+
+// https://developer.arm.com/documentation/ddi0596/2021-12/Base-Instructions/LDXR--Load-Exclusive-Register-
+// https://developer.arm.com/documentation/ddi0596/2021-12/Base-Instructions/LDXP--Load-Exclusive-Pair-of-Registers-
+// https://developer.arm.com/documentation/ddi0596/2021-12/Base-Instructions/STXR--Store-Exclusive-Register-
+// https://developer.arm.com/documentation/ddi0596/2021-12/Base-Instructions/STXP--Store-Exclusive-Pair-of-registers-
+union Exclusive {
+ constexpr explicit Exclusive(u32 raw_) : raw{raw_} {}
+
+ constexpr bool Verify() {
+ return this->GetSig() == 0x10;
+ }
+
+ constexpr u32 GetSig() {
+ return decltype(sig)::ExtractValue(raw);
+ }
+
+ constexpr u32 AsOrdered() {
+ return raw | decltype(o0)::FormatValue(1);
+ }
+
+ u32 raw;
+
+private:
+ BitField<0, 5, u32> rt; // memory operand
+ BitField<5, 5, u32> rn; // register operand 1
+ BitField<10, 5, u32> rt2; // register operand 2
+ BitField<15, 1, u32> o0; // ordered
+ BitField<16, 5, u32> rs; // status register
+ BitField<21, 2, u32> l; // operation type
+ BitField<23, 7, u32> sig; // 0x10
+ BitField<30, 2, u32> size; // size
+};
+static_assert(Exclusive(0xC85FFC00).Verify());
+static_assert(Exclusive(0xC85FFC00).AsOrdered() == 0xC85FFC00);
+static_assert(Exclusive(0xC85F7C00).AsOrdered() == 0xC85FFC00);
+static_assert(Exclusive(0xC8200440).AsOrdered() == 0xC8208440);
+
+} // namespace Core::NCE
diff --git a/src/core/arm/nce/patcher.cpp b/src/core/arm/nce/patcher.cpp
new file mode 100644
index 000000000..ec8527224
--- /dev/null
+++ b/src/core/arm/nce/patcher.cpp
@@ -0,0 +1,474 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "common/arm64/native_clock.h"
+#include "common/bit_cast.h"
+#include "common/literals.h"
+#include "core/arm/nce/arm_nce.h"
+#include "core/arm/nce/guest_context.h"
+#include "core/arm/nce/instructions.h"
+#include "core/arm/nce/patcher.h"
+#include "core/core.h"
+#include "core/core_timing.h"
+#include "core/hle/kernel/svc.h"
+
+namespace Core::NCE {
+
+using namespace Common::Literals;
+using namespace oaknut::util;
+
+using NativeExecutionParameters = Kernel::KThread::NativeExecutionParameters;
+
+constexpr size_t MaxRelativeBranch = 128_MiB;
+constexpr u32 ModuleCodeIndex = 0x24 / sizeof(u32);
+
+Patcher::Patcher() : c(m_patch_instructions) {}
+
+Patcher::~Patcher() = default;
+
+void Patcher::PatchText(const Kernel::PhysicalMemory& program_image,
+ const Kernel::CodeSet::Segment& code) {
+
+ // Write save context helper function.
+ c.l(m_save_context);
+ WriteSaveContext();
+
+ // Write load context helper function.
+ c.l(m_load_context);
+ WriteLoadContext();
+
+ // Retrieve text segment data.
+ const auto text = std::span{program_image}.subspan(code.offset, code.size);
+ const auto text_words =
+ std::span<const u32>{reinterpret_cast<const u32*>(text.data()), text.size() / sizeof(u32)};
+
+ // Loop through instructions, patching as needed.
+ for (u32 i = ModuleCodeIndex; i < static_cast<u32>(text_words.size()); i++) {
+ const u32 inst = text_words[i];
+
+ const auto AddRelocations = [&] {
+ const uintptr_t this_offset = i * sizeof(u32);
+ const uintptr_t next_offset = this_offset + sizeof(u32);
+
+ // Relocate from here to patch.
+ this->BranchToPatch(this_offset);
+
+ // Relocate from patch to next instruction.
+ return next_offset;
+ };
+
+ // SVC
+ if (auto svc = SVC{inst}; svc.Verify()) {
+ WriteSvcTrampoline(AddRelocations(), svc.GetValue());
+ continue;
+ }
+
+ // MRS Xn, TPIDR_EL0
+ // MRS Xn, TPIDRRO_EL0
+ if (auto mrs = MRS{inst};
+ mrs.Verify() && (mrs.GetSystemReg() == TpidrroEl0 || mrs.GetSystemReg() == TpidrEl0)) {
+ const auto src_reg = mrs.GetSystemReg() == TpidrroEl0 ? oaknut::SystemReg::TPIDRRO_EL0
+ : oaknut::SystemReg::TPIDR_EL0;
+ const auto dest_reg = oaknut::XReg{static_cast<int>(mrs.GetRt())};
+ WriteMrsHandler(AddRelocations(), dest_reg, src_reg);
+ continue;
+ }
+
+ // MRS Xn, CNTPCT_EL0
+ if (auto mrs = MRS{inst}; mrs.Verify() && mrs.GetSystemReg() == CntpctEl0) {
+ WriteCntpctHandler(AddRelocations(), oaknut::XReg{static_cast<int>(mrs.GetRt())});
+ continue;
+ }
+
+ // MRS Xn, CNTFRQ_EL0
+ if (auto mrs = MRS{inst}; mrs.Verify() && mrs.GetSystemReg() == CntfrqEl0) {
+ UNREACHABLE();
+ }
+
+ // MSR TPIDR_EL0, Xn
+ if (auto msr = MSR{inst}; msr.Verify() && msr.GetSystemReg() == TpidrEl0) {
+ WriteMsrHandler(AddRelocations(), oaknut::XReg{static_cast<int>(msr.GetRt())});
+ continue;
+ }
+
+ if (auto exclusive = Exclusive{inst}; exclusive.Verify()) {
+ m_exclusives.push_back(i);
+ }
+ }
+
+ // Determine patching mode for the final relocation step
+ const size_t image_size = program_image.size();
+ this->mode = image_size > MaxRelativeBranch ? PatchMode::PreText : PatchMode::PostData;
+}
+
+void Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
+ const Kernel::CodeSet::Segment& code,
+ Kernel::PhysicalMemory& program_image,
+ EntryTrampolines* out_trampolines) {
+ const size_t patch_size = GetSectionSize();
+ const size_t image_size = program_image.size();
+
+ // Retrieve text segment data.
+ const auto text = std::span{program_image}.subspan(code.offset, code.size);
+ const auto text_words =
+ std::span<u32>{reinterpret_cast<u32*>(text.data()), text.size() / sizeof(u32)};
+
+ const auto ApplyBranchToPatchRelocation = [&](u32* target, const Relocation& rel) {
+ oaknut::CodeGenerator rc{target};
+ if (mode == PatchMode::PreText) {
+ rc.B(rel.patch_offset - patch_size - rel.module_offset);
+ } else {
+ rc.B(image_size - rel.module_offset + rel.patch_offset);
+ }
+ };
+
+ const auto ApplyBranchToModuleRelocation = [&](u32* target, const Relocation& rel) {
+ oaknut::CodeGenerator rc{target};
+ if (mode == PatchMode::PreText) {
+ rc.B(patch_size - rel.patch_offset + rel.module_offset);
+ } else {
+ rc.B(rel.module_offset - image_size - rel.patch_offset);
+ }
+ };
+
+ const auto RebasePatch = [&](ptrdiff_t patch_offset) {
+ if (mode == PatchMode::PreText) {
+ return GetInteger(load_base) + patch_offset;
+ } else {
+ return GetInteger(load_base) + image_size + patch_offset;
+ }
+ };
+
+ const auto RebasePc = [&](uintptr_t module_offset) {
+ if (mode == PatchMode::PreText) {
+ return GetInteger(load_base) + patch_size + module_offset;
+ } else {
+ return GetInteger(load_base) + module_offset;
+ }
+ };
+
+ // We are now ready to relocate!
+ for (const Relocation& rel : m_branch_to_patch_relocations) {
+ ApplyBranchToPatchRelocation(text_words.data() + rel.module_offset / sizeof(u32), rel);
+ }
+ for (const Relocation& rel : m_branch_to_module_relocations) {
+ ApplyBranchToModuleRelocation(m_patch_instructions.data() + rel.patch_offset / sizeof(u32),
+ rel);
+ }
+
+ // Rewrite PC constants and record post trampolines
+ for (const Relocation& rel : m_write_module_pc_relocations) {
+ oaknut::CodeGenerator rc{m_patch_instructions.data() + rel.patch_offset / sizeof(u32)};
+ rc.dx(RebasePc(rel.module_offset));
+ }
+ for (const Trampoline& rel : m_trampolines) {
+ out_trampolines->insert({RebasePc(rel.module_offset), RebasePatch(rel.patch_offset)});
+ }
+
+ // Cortex-A57 seems to treat all exclusives as ordered, but newer processors do not.
+ // Convert to ordered to preserve this assumption.
+ for (const ModuleTextAddress i : m_exclusives) {
+ auto exclusive = Exclusive{text_words[i]};
+ text_words[i] = exclusive.AsOrdered();
+ }
+
+ // Copy to program image
+ if (this->mode == PatchMode::PreText) {
+ std::memcpy(program_image.data(), m_patch_instructions.data(),
+ m_patch_instructions.size() * sizeof(u32));
+ } else {
+ program_image.resize(image_size + patch_size);
+ std::memcpy(program_image.data() + image_size, m_patch_instructions.data(),
+ m_patch_instructions.size() * sizeof(u32));
+ }
+}
+
+size_t Patcher::GetSectionSize() const noexcept {
+ return Common::AlignUp(m_patch_instructions.size() * sizeof(u32), Core::Memory::YUZU_PAGESIZE);
+}
+
+void Patcher::WriteLoadContext() {
+ // This function was called, which modifies X30, so use that as a scratch register.
+ // SP contains the guest X30, so save our return X30 to SP + 8, since we have allocated 16 bytes
+ // of stack.
+ c.STR(X30, SP, 8);
+ c.MRS(X30, oaknut::SystemReg::TPIDR_EL0);
+ c.LDR(X30, X30, offsetof(NativeExecutionParameters, native_context));
+
+ // Load system registers.
+ c.LDR(W0, X30, offsetof(GuestContext, fpsr));
+ c.MSR(oaknut::SystemReg::FPSR, X0);
+ c.LDR(W0, X30, offsetof(GuestContext, fpcr));
+ c.MSR(oaknut::SystemReg::FPCR, X0);
+ c.LDR(W0, X30, offsetof(GuestContext, nzcv));
+ c.MSR(oaknut::SystemReg::NZCV, X0);
+
+ // Load all vector registers.
+ static constexpr size_t VEC_OFF = offsetof(GuestContext, vector_registers);
+ for (int i = 0; i <= 30; i += 2) {
+ c.LDP(oaknut::QReg{i}, oaknut::QReg{i + 1}, X30, VEC_OFF + 16 * i);
+ }
+
+ // Load all general-purpose registers except X30.
+ for (int i = 0; i <= 28; i += 2) {
+ c.LDP(oaknut::XReg{i}, oaknut::XReg{i + 1}, X30, 8 * i);
+ }
+
+ // Reload our return X30 from the stack and return.
+ // The patch code will reload the guest X30 for us.
+ c.LDR(X30, SP, 8);
+ c.RET();
+}
+
+void Patcher::WriteSaveContext() {
+ // This function was called, which modifies X30, so use that as a scratch register.
+ // SP contains the guest X30, so save our X30 to SP + 8, since we have allocated 16 bytes of
+ // stack.
+ c.STR(X30, SP, 8);
+ c.MRS(X30, oaknut::SystemReg::TPIDR_EL0);
+ c.LDR(X30, X30, offsetof(NativeExecutionParameters, native_context));
+
+ // Store all general-purpose registers except X30.
+ for (int i = 0; i <= 28; i += 2) {
+ c.STP(oaknut::XReg{i}, oaknut::XReg{i + 1}, X30, 8 * i);
+ }
+
+ // Store all vector registers.
+ static constexpr size_t VEC_OFF = offsetof(GuestContext, vector_registers);
+ for (int i = 0; i <= 30; i += 2) {
+ c.STP(oaknut::QReg{i}, oaknut::QReg{i + 1}, X30, VEC_OFF + 16 * i);
+ }
+
+ // Store guest system registers, X30 and SP, using X0 as a scratch register.
+ c.STR(X0, SP, PRE_INDEXED, -16);
+ c.LDR(X0, SP, 16);
+ c.STR(X0, X30, 8 * 30);
+ c.ADD(X0, SP, 32);
+ c.STR(X0, X30, offsetof(GuestContext, sp));
+ c.MRS(X0, oaknut::SystemReg::FPSR);
+ c.STR(W0, X30, offsetof(GuestContext, fpsr));
+ c.MRS(X0, oaknut::SystemReg::FPCR);
+ c.STR(W0, X30, offsetof(GuestContext, fpcr));
+ c.MRS(X0, oaknut::SystemReg::NZCV);
+ c.STR(W0, X30, offsetof(GuestContext, nzcv));
+ c.LDR(X0, SP, POST_INDEXED, 16);
+
+ // Reload our return X30 from the stack, and return.
+ c.LDR(X30, SP, 8);
+ c.RET();
+}
+
+void Patcher::WriteSvcTrampoline(ModuleDestLabel module_dest, u32 svc_id) {
+ // We are about to start saving state, so we need to lock the context.
+ this->LockContext();
+
+ // Store guest X30 to the stack. Then, save the context and restore the stack.
+ // This will save all registers except PC, but we know PC at patch time.
+ c.STR(X30, SP, PRE_INDEXED, -16);
+ c.BL(m_save_context);
+ c.LDR(X30, SP, POST_INDEXED, 16);
+
+ // Now that we've saved all registers, we can use any registers as scratch.
+ // Store PC + 4 to arm interface, since we know the instruction offset from the entry point.
+ oaknut::Label pc_after_svc;
+ c.MRS(X1, oaknut::SystemReg::TPIDR_EL0);
+ c.LDR(X1, X1, offsetof(NativeExecutionParameters, native_context));
+ c.LDR(X2, pc_after_svc);
+ c.STR(X2, X1, offsetof(GuestContext, pc));
+
+ // Store SVC number to execute when we return
+ c.MOV(X2, svc_id);
+ c.STR(W2, X1, offsetof(GuestContext, svc_swi));
+
+ // We are calling a SVC. Clear esr_el1 and return it.
+ static_assert(std::is_same_v<std::underlying_type_t<HaltReason>, u64>);
+ oaknut::Label retry;
+ c.ADD(X2, X1, offsetof(GuestContext, esr_el1));
+ c.l(retry);
+ c.LDAXR(X0, X2);
+ c.STLXR(W3, XZR, X2);
+ c.CBNZ(W3, retry);
+
+ // Add "calling SVC" flag. Since this is X0, this is now our return value.
+ c.ORR(X0, X0, static_cast<u64>(HaltReason::SupervisorCall));
+
+ // Offset the GuestContext pointer to the HostContext member.
+ // STP has limited range of [-512, 504] which we can't reach otherwise
+ // NB: Due to this all offsets below are from the start of HostContext.
+ c.ADD(X1, X1, offsetof(GuestContext, host_ctx));
+
+ // Reload host TPIDR_EL0 and SP.
+ static_assert(offsetof(HostContext, host_sp) + 8 == offsetof(HostContext, host_tpidr_el0));
+ c.LDP(X2, X3, X1, offsetof(HostContext, host_sp));
+ c.MOV(SP, X2);
+ c.MSR(oaknut::SystemReg::TPIDR_EL0, X3);
+
+ // Load callee-saved host registers and return to host.
+ static constexpr size_t HOST_REGS_OFF = offsetof(HostContext, host_saved_regs);
+ static constexpr size_t HOST_VREGS_OFF = offsetof(HostContext, host_saved_vregs);
+ c.LDP(X19, X20, X1, HOST_REGS_OFF);
+ c.LDP(X21, X22, X1, HOST_REGS_OFF + 2 * sizeof(u64));
+ c.LDP(X23, X24, X1, HOST_REGS_OFF + 4 * sizeof(u64));
+ c.LDP(X25, X26, X1, HOST_REGS_OFF + 6 * sizeof(u64));
+ c.LDP(X27, X28, X1, HOST_REGS_OFF + 8 * sizeof(u64));
+ c.LDP(X29, X30, X1, HOST_REGS_OFF + 10 * sizeof(u64));
+ c.LDP(Q8, Q9, X1, HOST_VREGS_OFF);
+ c.LDP(Q10, Q11, X1, HOST_VREGS_OFF + 2 * sizeof(u128));
+ c.LDP(Q12, Q13, X1, HOST_VREGS_OFF + 4 * sizeof(u128));
+ c.LDP(Q14, Q15, X1, HOST_VREGS_OFF + 6 * sizeof(u128));
+ c.RET();
+
+ // Write the post-SVC trampoline address, which will jump back to the guest after restoring its
+ // state.
+ m_trampolines.push_back({c.offset(), module_dest});
+
+ // Host called this location. Save the return address so we can
+ // unwind the stack properly when jumping back.
+ c.MRS(X2, oaknut::SystemReg::TPIDR_EL0);
+ c.LDR(X2, X2, offsetof(NativeExecutionParameters, native_context));
+ c.ADD(X0, X2, offsetof(GuestContext, host_ctx));
+ c.STR(X30, X0, offsetof(HostContext, host_saved_regs) + 11 * sizeof(u64));
+
+ // Reload all guest registers except X30 and PC.
+ // The function also expects 16 bytes of stack already allocated.
+ c.STR(X30, SP, PRE_INDEXED, -16);
+ c.BL(m_load_context);
+ c.LDR(X30, SP, POST_INDEXED, 16);
+
+ // Use X1 as a scratch register to restore X30.
+ c.STR(X1, SP, PRE_INDEXED, -16);
+ c.MRS(X1, oaknut::SystemReg::TPIDR_EL0);
+ c.LDR(X1, X1, offsetof(NativeExecutionParameters, native_context));
+ c.LDR(X30, X1, offsetof(GuestContext, cpu_registers) + sizeof(u64) * 30);
+ c.LDR(X1, SP, POST_INDEXED, 16);
+
+ // Unlock the context.
+ this->UnlockContext();
+
+ // Jump back to the instruction after the emulated SVC.
+ this->BranchToModule(module_dest);
+
+ // Store PC after call.
+ c.l(pc_after_svc);
+ this->WriteModulePc(module_dest);
+}
+
+void Patcher::WriteMrsHandler(ModuleDestLabel module_dest, oaknut::XReg dest_reg,
+ oaknut::SystemReg src_reg) {
+ // Retrieve emulated TLS register from GuestContext.
+ c.MRS(dest_reg, oaknut::SystemReg::TPIDR_EL0);
+ if (src_reg == oaknut::SystemReg::TPIDRRO_EL0) {
+ c.LDR(dest_reg, dest_reg, offsetof(NativeExecutionParameters, tpidrro_el0));
+ } else {
+ c.LDR(dest_reg, dest_reg, offsetof(NativeExecutionParameters, tpidr_el0));
+ }
+
+ // Jump back to the instruction after the emulated MRS.
+ this->BranchToModule(module_dest);
+}
+
+void Patcher::WriteMsrHandler(ModuleDestLabel module_dest, oaknut::XReg src_reg) {
+ const auto scratch_reg = src_reg.index() == 0 ? X1 : X0;
+ c.STR(scratch_reg, SP, PRE_INDEXED, -16);
+
+ // Save guest value to NativeExecutionParameters::tpidr_el0.
+ c.MRS(scratch_reg, oaknut::SystemReg::TPIDR_EL0);
+ c.STR(src_reg, scratch_reg, offsetof(NativeExecutionParameters, tpidr_el0));
+
+ // Restore scratch register.
+ c.LDR(scratch_reg, SP, POST_INDEXED, 16);
+
+ // Jump back to the instruction after the emulated MSR.
+ this->BranchToModule(module_dest);
+}
+
+void Patcher::WriteCntpctHandler(ModuleDestLabel module_dest, oaknut::XReg dest_reg) {
+ static Common::Arm64::NativeClock clock{};
+ const auto factor = clock.GetGuestCNTFRQFactor();
+ const auto raw_factor = Common::BitCast<std::array<u64, 2>>(factor);
+
+ const auto use_x2_x3 = dest_reg.index() == 0 || dest_reg.index() == 1;
+ oaknut::XReg scratch0 = use_x2_x3 ? X2 : X0;
+ oaknut::XReg scratch1 = use_x2_x3 ? X3 : X1;
+
+ oaknut::Label factorlo;
+ oaknut::Label factorhi;
+
+ // Save scratches.
+ c.STP(scratch0, scratch1, SP, PRE_INDEXED, -16);
+
+ // Load counter value.
+ c.MRS(dest_reg, oaknut::SystemReg::CNTVCT_EL0);
+
+ // Load scaling factor.
+ c.LDR(scratch0, factorlo);
+ c.LDR(scratch1, factorhi);
+
+ // Multiply low bits and get result.
+ c.UMULH(scratch0, dest_reg, scratch0);
+
+ // Multiply high bits and add low bit result.
+ c.MADD(dest_reg, dest_reg, scratch1, scratch0);
+
+ // Reload scratches.
+ c.LDP(scratch0, scratch1, SP, POST_INDEXED, 16);
+
+ // Jump back to the instruction after the emulated MRS.
+ this->BranchToModule(module_dest);
+
+ // Scaling factor constant values.
+ c.l(factorlo);
+ c.dx(raw_factor[0]);
+ c.l(factorhi);
+ c.dx(raw_factor[1]);
+}
+
+void Patcher::LockContext() {
+ oaknut::Label retry;
+
+ // Save scratches.
+ c.STP(X0, X1, SP, PRE_INDEXED, -16);
+
+ // Reload lock pointer.
+ c.l(retry);
+ c.CLREX();
+ c.MRS(X0, oaknut::SystemReg::TPIDR_EL0);
+ c.ADD(X0, X0, offsetof(NativeExecutionParameters, lock));
+
+ static_assert(SpinLockLocked == 0);
+
+ // Load-linked with acquire ordering.
+ c.LDAXR(W1, X0);
+
+ // If the value was SpinLockLocked, clear monitor and retry.
+ c.CBZ(W1, retry);
+
+ // Store-conditional SpinLockLocked with relaxed ordering.
+ c.STXR(W1, WZR, X0);
+
+ // If we failed to store, retry.
+ c.CBNZ(W1, retry);
+
+ // We succeeded! Reload scratches.
+ c.LDP(X0, X1, SP, POST_INDEXED, 16);
+}
+
+void Patcher::UnlockContext() {
+ // Save scratches.
+ c.STP(X0, X1, SP, PRE_INDEXED, -16);
+
+ // Load lock pointer.
+ c.MRS(X0, oaknut::SystemReg::TPIDR_EL0);
+ c.ADD(X0, X0, offsetof(NativeExecutionParameters, lock));
+
+ // Load SpinLockUnlocked.
+ c.MOV(W1, SpinLockUnlocked);
+
+ // Store value with release ordering.
+ c.STLR(W1, X0);
+
+ // Load scratches.
+ c.LDP(X0, X1, SP, POST_INDEXED, 16);
+}
+
+} // namespace Core::NCE
diff --git a/src/core/arm/nce/patcher.h b/src/core/arm/nce/patcher.h
new file mode 100644
index 000000000..c6d1608c1
--- /dev/null
+++ b/src/core/arm/nce/patcher.h
@@ -0,0 +1,98 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <span>
+#include <unordered_map>
+#include <vector>
+#include <oaknut/code_block.hpp>
+#include <oaknut/oaknut.hpp>
+
+#include "common/common_types.h"
+#include "core/hle/kernel/code_set.h"
+#include "core/hle/kernel/k_typed_address.h"
+#include "core/hle/kernel/physical_memory.h"
+
+namespace Core::NCE {
+
+enum class PatchMode : u32 {
+ None,
+ PreText, ///< Patch section is inserted before .text
+ PostData, ///< Patch section is inserted after .data
+};
+
+using ModuleTextAddress = u64;
+using PatchTextAddress = u64;
+using EntryTrampolines = std::unordered_map<ModuleTextAddress, PatchTextAddress>;
+
+class Patcher {
+public:
+ explicit Patcher();
+ ~Patcher();
+
+ void PatchText(const Kernel::PhysicalMemory& program_image,
+ const Kernel::CodeSet::Segment& code);
+ void RelocateAndCopy(Common::ProcessAddress load_base, const Kernel::CodeSet::Segment& code,
+ Kernel::PhysicalMemory& program_image, EntryTrampolines* out_trampolines);
+ size_t GetSectionSize() const noexcept;
+
+ [[nodiscard]] PatchMode GetPatchMode() const noexcept {
+ return mode;
+ }
+
+private:
+ using ModuleDestLabel = uintptr_t;
+
+ struct Trampoline {
+ ptrdiff_t patch_offset;
+ uintptr_t module_offset;
+ };
+
+ void WriteLoadContext();
+ void WriteSaveContext();
+ void LockContext();
+ void UnlockContext();
+ void WriteSvcTrampoline(ModuleDestLabel module_dest, u32 svc_id);
+ void WriteMrsHandler(ModuleDestLabel module_dest, oaknut::XReg dest_reg,
+ oaknut::SystemReg src_reg);
+ void WriteMsrHandler(ModuleDestLabel module_dest, oaknut::XReg src_reg);
+ void WriteCntpctHandler(ModuleDestLabel module_dest, oaknut::XReg dest_reg);
+
+private:
+ void BranchToPatch(uintptr_t module_dest) {
+ m_branch_to_patch_relocations.push_back({c.offset(), module_dest});
+ }
+
+ void BranchToModule(uintptr_t module_dest) {
+ m_branch_to_module_relocations.push_back({c.offset(), module_dest});
+ c.dw(0);
+ }
+
+ void WriteModulePc(uintptr_t module_dest) {
+ m_write_module_pc_relocations.push_back({c.offset(), module_dest});
+ c.dx(0);
+ }
+
+private:
+ // List of patch instructions we have generated.
+ std::vector<u32> m_patch_instructions{};
+
+ // Relocation type for relative branch from module to patch.
+ struct Relocation {
+ ptrdiff_t patch_offset; ///< Offset in bytes from the start of the patch section.
+ uintptr_t module_offset; ///< Offset in bytes from the start of the text section.
+ };
+
+ oaknut::VectorCodeGenerator c;
+ std::vector<Trampoline> m_trampolines;
+ std::vector<Relocation> m_branch_to_patch_relocations{};
+ std::vector<Relocation> m_branch_to_module_relocations{};
+ std::vector<Relocation> m_write_module_pc_relocations{};
+ std::vector<ModuleTextAddress> m_exclusives{};
+ oaknut::Label m_save_context{};
+ oaknut::Label m_load_context{};
+ PatchMode mode{PatchMode::None};
+};
+
+} // namespace Core::NCE