early-access version 4000

README.md

@@ -1,7 +1,7 @@
 yuzu emulator early access
 =============
 
-This is the source code for early-access 3999.
+This is the source code for early-access 4000.
 
 ## Legal Notice
 

externals/CMakeLists.txt

@@ -193,7 +193,7 @@ if (ANDROID)
    endif()
 endif()
 
-if (CMAKE_SYSTEM_NAME STREQUAL "Linux")
+if (UNIX AND NOT APPLE)
     add_subdirectory(gamemode)
 endif()
 

src/common/CMakeLists.txt

@@ -182,7 +182,7 @@ if(ANDROID)
     )
 endif()
 
-if (CMAKE_SYSTEM_NAME STREQUAL "Linux")
+if (UNIX AND NOT APPLE)
   target_sources(common PRIVATE
     linux/gamemode.cpp
     linux/gamemode.h

src/core/CMakeLists.txt

@@ -935,8 +935,8 @@ if (HAS_NCE)
         arm/nce/arm_nce.h
         arm/nce/arm_nce.s
         arm/nce/guest_context.h
-        arm/nce/patch.cpp
+        arm/nce/patcher.cpp
-        arm/nce/patch.h
+        arm/nce/patcher.h
         arm/nce/instructions.h
     )
     target_link_libraries(core PRIVATE merry::oaknut)

src/core/arm/nce/arm_nce.cpp

@@ -6,7 +6,7 @@
 
 #include "common/signal_chain.h"
 #include "core/arm/nce/arm_nce.h"
-#include "core/arm/nce/patch.h"
+#include "core/arm/nce/patcher.h"
 #include "core/core.h"
 #include "core/memory.h"
 

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

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

src/core/loader/deconstructed_rom_directory.cpp

@@ -16,7 +16,7 @@
 #include "core/loader/nso.h"
 
 #ifdef HAS_NCE
-#include "core/arm/nce/patch.h"
+#include "core/arm/nce/patcher.h"
 #endif
 
 namespace Loader {

src/core/loader/nro.cpp

@@ -23,7 +23,7 @@
 #include "core/memory.h"
 
 #ifdef HAS_NCE
-#include "core/arm/nce/patch.h"
+#include "core/arm/nce/patcher.h"
 #endif
 
 namespace Loader {

src/core/loader/nso.cpp

@@ -21,7 +21,7 @@
 #include "core/memory.h"
 
 #ifdef HAS_NCE
-#include "core/arm/nce/patch.h"
+#include "core/arm/nce/patcher.h"
 #endif
 
 namespace Loader {

src/video_core/renderer_opengl/gl_buffer_cache.cpp

@@ -146,8 +146,12 @@ StagingBufferMap BufferCacheRuntime::UploadStagingBuffer(size_t size) {
     return staging_buffer_pool.RequestUploadBuffer(size);
 }
 
-StagingBufferMap BufferCacheRuntime::DownloadStagingBuffer(size_t size) {
+StagingBufferMap BufferCacheRuntime::DownloadStagingBuffer(size_t size, bool deferred) {
-    return staging_buffer_pool.RequestDownloadBuffer(size);
+    return staging_buffer_pool.RequestDownloadBuffer(size, deferred);
+}
+
+void BufferCacheRuntime::FreeDeferredStagingBuffer(StagingBufferMap& buffer) {
+    staging_buffer_pool.FreeDeferredStagingBuffer(buffer);
 }
 
 u64 BufferCacheRuntime::GetDeviceMemoryUsage() const {

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -66,7 +66,9 @@ public:
 
     [[nodiscard]] StagingBufferMap UploadStagingBuffer(size_t size);
 
-    [[nodiscard]] StagingBufferMap DownloadStagingBuffer(size_t size);
+    [[nodiscard]] StagingBufferMap DownloadStagingBuffer(size_t size, bool deferred = false);
+
+    void FreeDeferredStagingBuffer(StagingBufferMap& buffer);
 
     bool CanReorderUpload(const Buffer&, std::span<const VideoCommon::BufferCopy>) {
         return false;
@@ -246,7 +248,7 @@ struct BufferCacheParams {
     static constexpr bool NEEDS_BIND_STORAGE_INDEX = true;
     static constexpr bool USE_MEMORY_MAPS = true;
     static constexpr bool SEPARATE_IMAGE_BUFFER_BINDINGS = true;
-    static constexpr bool IMPLEMENTS_ASYNC_DOWNLOADS = false;
+    static constexpr bool IMPLEMENTS_ASYNC_DOWNLOADS = true;
 
     // TODO: Investigate why OpenGL seems to perform worse with persistently mapped buffer uploads
     static constexpr bool USE_MEMORY_MAPS_FOR_UPLOADS = false;

src/video_core/renderer_opengl/gl_staging_buffer_pool.cpp

@@ -28,63 +28,69 @@ StagingBuffers::StagingBuffers(GLenum storage_flags_, GLenum map_flags_)
 
 StagingBuffers::~StagingBuffers() = default;
 
-StagingBufferMap StagingBuffers::RequestMap(size_t requested_size, bool insert_fence) {
+StagingBufferMap StagingBuffers::RequestMap(size_t requested_size, bool insert_fence,
+                                            bool deferred) {
     MICROPROFILE_SCOPE(OpenGL_BufferRequest);
 
     const size_t index = RequestBuffer(requested_size);
-    OGLSync* const sync = insert_fence ? &syncs[index] : nullptr;
+    OGLSync* const sync = insert_fence ? &allocs[index].sync : nullptr;
-    sync_indices[index] = insert_fence ? ++current_sync_index : 0;
+    allocs[index].sync_index = insert_fence ? ++current_sync_index : 0;
+    allocs[index].deferred = deferred;
     return StagingBufferMap{
-        .mapped_span = std::span(maps[index], requested_size),
+        .mapped_span = std::span(allocs[index].map, requested_size),
         .sync = sync,
-        .buffer = buffers[index].handle,
+        .buffer = allocs[index].buffer.handle,
+        .index = index,
     };
 }
 
+void StagingBuffers::FreeDeferredStagingBuffer(size_t index) {
+    ASSERT(allocs[index].deferred);
+    allocs[index].deferred = false;
+}
+
 size_t StagingBuffers::RequestBuffer(size_t requested_size) {
     if (const std::optional<size_t> index = FindBuffer(requested_size); index) {
         return *index;
     }
-
+    StagingBufferAlloc alloc;
-    OGLBuffer& buffer = buffers.emplace_back();
+    alloc.buffer.Create();
-    buffer.Create();
     const auto next_pow2_size = Common::NextPow2(requested_size);
-    glNamedBufferStorage(buffer.handle, next_pow2_size, nullptr,
+    glNamedBufferStorage(alloc.buffer.handle, next_pow2_size, nullptr,
                          storage_flags | GL_MAP_PERSISTENT_BIT);
-    maps.push_back(static_cast<u8*>(glMapNamedBufferRange(buffer.handle, 0, next_pow2_size,
+    alloc.map = static_cast<u8*>(glMapNamedBufferRange(alloc.buffer.handle, 0, next_pow2_size,
-                                                          map_flags | GL_MAP_PERSISTENT_BIT)));
+                                                       map_flags | GL_MAP_PERSISTENT_BIT));
-    syncs.emplace_back();
+    alloc.size = next_pow2_size;
-    sync_indices.emplace_back();
+    allocs.emplace_back(std::move(alloc));
-    sizes.push_back(next_pow2_size);
+    return allocs.size() - 1;
-
-    ASSERT(syncs.size() == buffers.size() && buffers.size() == maps.size() &&
-           maps.size() == sizes.size());
-
-    return buffers.size() - 1;
 }
 
 std::optional<size_t> StagingBuffers::FindBuffer(size_t requested_size) {
     size_t known_unsignaled_index = current_sync_index + 1;
     size_t smallest_buffer = std::numeric_limits<size_t>::max();
     std::optional<size_t> found;
-    const size_t num_buffers = sizes.size();
+    const size_t num_buffers = allocs.size();
     for (size_t index = 0; index < num_buffers; ++index) {
-        const size_t buffer_size = sizes[index];
+        StagingBufferAlloc& alloc = allocs[index];
+        const size_t buffer_size = alloc.size;
         if (buffer_size < requested_size || buffer_size >= smallest_buffer) {
             continue;
         }
-        if (syncs[index].handle != 0) {
+        if (alloc.deferred) {
-            if (sync_indices[index] >= known_unsignaled_index) {
+            continue;
+        }
+        if (alloc.sync.handle != 0) {
+            if (alloc.sync_index >= known_unsignaled_index) {
                 // This fence is later than a fence that is known to not be signaled
                 continue;
             }
-            if (!syncs[index].IsSignaled()) {
+            if (!alloc.sync.IsSignaled()) {
                 // Since this fence hasn't been signaled, it's safe to assume all later
                 // fences haven't been signaled either
-                known_unsignaled_index = std::min(known_unsignaled_index, sync_indices[index]);
+                known_unsignaled_index = std::min(known_unsignaled_index, alloc.sync_index);
                 continue;
             }
-            syncs[index].Release();
+            alloc.sync.Release();
         }
         smallest_buffer = buffer_size;
         found = index;
@@ -143,8 +149,12 @@ StagingBufferMap StagingBufferPool::RequestUploadBuffer(size_t size) {
     return upload_buffers.RequestMap(size, true);
 }
 
-StagingBufferMap StagingBufferPool::RequestDownloadBuffer(size_t size) {
+StagingBufferMap StagingBufferPool::RequestDownloadBuffer(size_t size, bool deferred) {
-    return download_buffers.RequestMap(size, false);
+    return download_buffers.RequestMap(size, false, deferred);
+}
+
+void StagingBufferPool::FreeDeferredStagingBuffer(StagingBufferMap& buffer) {
+    download_buffers.FreeDeferredStagingBuffer(buffer.index);
 }
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_staging_buffer_pool.h

@@ -26,23 +26,30 @@ struct StagingBufferMap {
     size_t offset = 0;
     OGLSync* sync;
     GLuint buffer;
+    size_t index;
 };
 
 struct StagingBuffers {
     explicit StagingBuffers(GLenum storage_flags_, GLenum map_flags_);
     ~StagingBuffers();
 
-    StagingBufferMap RequestMap(size_t requested_size, bool insert_fence);
+    StagingBufferMap RequestMap(size_t requested_size, bool insert_fence, bool deferred = false);
+
+    void FreeDeferredStagingBuffer(size_t index);
 
     size_t RequestBuffer(size_t requested_size);
 
     std::optional<size_t> FindBuffer(size_t requested_size);
 
-    std::vector<OGLSync> syncs;
+    struct StagingBufferAlloc {
-    std::vector<OGLBuffer> buffers;
+        OGLSync sync;
-    std::vector<u8*> maps;
+        OGLBuffer buffer;
-    std::vector<size_t> sizes;
+        u8* map;
-    std::vector<size_t> sync_indices;
+        size_t size;
+        size_t sync_index;
+        bool deferred;
+    };
+    std::vector<StagingBufferAlloc> allocs;
     GLenum storage_flags;
     GLenum map_flags;
     size_t current_sync_index = 0;
@@ -85,7 +92,8 @@ public:
     ~StagingBufferPool() = default;
 
     StagingBufferMap RequestUploadBuffer(size_t size);
-    StagingBufferMap RequestDownloadBuffer(size_t size);
+    StagingBufferMap RequestDownloadBuffer(size_t size, bool deferred = false);
+    void FreeDeferredStagingBuffer(StagingBufferMap& buffer);
 
 private:
     StagingBuffers upload_buffers{GL_MAP_WRITE_BIT, GL_MAP_WRITE_BIT | GL_MAP_FLUSH_EXPLICIT_BIT};

src/video_core/renderer_opengl/gl_texture_cache.cpp

@@ -557,8 +557,12 @@ StagingBufferMap TextureCacheRuntime::UploadStagingBuffer(size_t size) {
     return staging_buffer_pool.RequestUploadBuffer(size);
 }
 
-StagingBufferMap TextureCacheRuntime::DownloadStagingBuffer(size_t size) {
+StagingBufferMap TextureCacheRuntime::DownloadStagingBuffer(size_t size, bool deferred) {
-    return staging_buffer_pool.RequestDownloadBuffer(size);
+    return staging_buffer_pool.RequestDownloadBuffer(size, deferred);
+}
+
+void TextureCacheRuntime::FreeDeferredStagingBuffer(StagingBufferMap& buffer) {
+    staging_buffer_pool.FreeDeferredStagingBuffer(buffer);
 }
 
 u64 TextureCacheRuntime::GetDeviceMemoryUsage() const {

src/video_core/renderer_opengl/gl_texture_cache.h

@@ -74,7 +74,9 @@ public:
 
     StagingBufferMap UploadStagingBuffer(size_t size);
 
-    StagingBufferMap DownloadStagingBuffer(size_t size);
+    StagingBufferMap DownloadStagingBuffer(size_t size, bool deferred = false);
+
+    void FreeDeferredStagingBuffer(StagingBufferMap& buffer);
 
     u64 GetDeviceLocalMemory() const {
         return device_access_memory;
@@ -359,7 +361,7 @@ struct TextureCacheParams {
     static constexpr bool FRAMEBUFFER_BLITS = true;
     static constexpr bool HAS_EMULATED_COPIES = true;
     static constexpr bool HAS_DEVICE_MEMORY_INFO = true;
-    static constexpr bool IMPLEMENTS_ASYNC_DOWNLOADS = false;
+    static constexpr bool IMPLEMENTS_ASYNC_DOWNLOADS = true;
 
     using Runtime = OpenGL::TextureCacheRuntime;
     using Image = OpenGL::Image;
@@ -367,7 +369,7 @@ struct TextureCacheParams {
     using ImageView = OpenGL::ImageView;
     using Sampler = OpenGL::Sampler;
     using Framebuffer = OpenGL::Framebuffer;
-    using AsyncBuffer = u32;
+    using AsyncBuffer = OpenGL::StagingBufferMap;
     using BufferType = GLuint;
 };
 

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -892,10 +892,6 @@ void RasterizerVulkan::UpdateDynamicStates() {
         UpdateFrontFace(regs);
         UpdateStencilOp(regs);
 
-        if (device.IsExtVertexInputDynamicStateSupported()) {
-            UpdateVertexInput(regs);
-        }
-
         if (state_tracker.TouchStateEnable()) {
             UpdateDepthBoundsTestEnable(regs);
             UpdateDepthTestEnable(regs);
@@ -918,6 +914,9 @@ void RasterizerVulkan::UpdateDynamicStates() {
             UpdateBlending(regs);
         }
     }
+    if (device.IsExtVertexInputDynamicStateSupported()) {
+        UpdateVertexInput(regs);
+    }
 }
 
 void RasterizerVulkan::HandleTransformFeedback() {

src/video_core/vulkan_common/vulkan_device.cpp

@@ -519,10 +519,6 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
         LOG_WARNING(Render_Vulkan, "ARM drivers have broken VK_EXT_extended_dynamic_state");
         RemoveExtensionFeature(extensions.extended_dynamic_state, features.extended_dynamic_state,
                                VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME);
-
-        LOG_WARNING(Render_Vulkan, "ARM drivers have broken VK_EXT_extended_dynamic_state2");
-        RemoveExtensionFeature(extensions.extended_dynamic_state2, features.extended_dynamic_state2,
-                               VK_EXT_EXTENDED_DYNAMIC_STATE_2_EXTENSION_NAME);
     }
 
     if (is_nvidia) {
@@ -611,17 +607,12 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
         }
     }
     if (extensions.vertex_input_dynamic_state && is_qualcomm) {
-        const u32 version = (properties.properties.driverVersion << 3) >> 3;
+        // Qualcomm drivers do not properly support vertex_input_dynamic_state.
-        if (version >= VK_MAKE_API_VERSION(0, 0, 676, 0) &&
+        LOG_WARNING(Render_Vulkan,
-            version < VK_MAKE_API_VERSION(0, 0, 680, 0)) {
+                    "Qualcomm drivers have broken VK_EXT_vertex_input_dynamic_state");
-            // Qualcomm Adreno 7xx drivers do not properly support vertex_input_dynamic_state.
+        RemoveExtensionFeature(extensions.vertex_input_dynamic_state,
-            LOG_WARNING(
+                               features.vertex_input_dynamic_state,
-                Render_Vulkan,
+                               VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME);
-                "Qualcomm Adreno 7xx drivers have broken VK_EXT_vertex_input_dynamic_state");
-            RemoveExtensionFeature(extensions.vertex_input_dynamic_state,
-                                   features.vertex_input_dynamic_state,
-                                   VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME);
-        }
     }
 
     sets_per_pool = 64;
@@ -704,6 +695,22 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
             std::min(properties.properties.limits.maxVertexInputBindings, 16U);
     }
 
+    if (!extensions.extended_dynamic_state && extensions.extended_dynamic_state2) {
+        LOG_INFO(Render_Vulkan,
+                 "Removing extendedDynamicState2 due to missing extendedDynamicState");
+        RemoveExtensionFeature(extensions.extended_dynamic_state2, features.extended_dynamic_state2,
+                               VK_EXT_EXTENDED_DYNAMIC_STATE_2_EXTENSION_NAME);
+    }
+
+    if (!extensions.extended_dynamic_state2 && extensions.extended_dynamic_state3) {
+        LOG_INFO(Render_Vulkan,
+                 "Removing extendedDynamicState3 due to missing extendedDynamicState2");
+        RemoveExtensionFeature(extensions.extended_dynamic_state3, features.extended_dynamic_state3,
+                               VK_EXT_EXTENDED_DYNAMIC_STATE_3_EXTENSION_NAME);
+        dynamic_state3_blending = false;
+        dynamic_state3_enables = false;
+    }
+
     logical = vk::Device::Create(physical, queue_cis, ExtensionListForVulkan(loaded_extensions),
                                  first_next, dld);