[dynarmic, loongarch64] Add minimal toy implementation enough to execute LSLS (#4054)

Reviewed-on: https://git.eden-emu.dev/eden-emu/eden/pulls/4054
Reviewed-by: Lizzie <lizzie@eden-emu.dev>
Reviewed-by: crueter <crueter@eden-emu.dev>

src/dynarmic/src/dynarmic/CMakeLists.txt

@@ -299,10 +299,24 @@ if ("loongarch64" IN_LIST ARCHITECTURE)
     target_link_libraries(dynarmic PRIVATE lagoon::lagoon)
 
     target_sources(dynarmic PRIVATE
-        backend/loongarch64/exclusive_monitor.cpp
+        backend/loongarch64/abi.h
+        backend/loongarch64/a32_jitstate.cpp
+        backend/loongarch64/a32_jitstate.h
+        backend/loongarch64/code_block.h
+        backend/loongarch64/emit_context.h
+        backend/loongarch64/emit_loongarch64.cpp
+        backend/loongarch64/emit_loongarch64.h
+        backend/loongarch64/emit_loongarch64_a32.cpp
+        backend/loongarch64/emit_loongarch64_data_processing.cpp
+        backend/loongarch64/a32_address_space.cpp
+        backend/loongarch64/a32_address_space.h
         backend/loongarch64/a32_interface.cpp
         backend/loongarch64/a64_interface.cpp
-        backend/loongarch64/code_block.h
+        backend/loongarch64/exclusive_monitor.cpp
+        backend/loongarch64/reg_alloc.cpp
+        backend/loongarch64/reg_alloc.h
+        backend/loongarch64/stack_layout.h
+        backend/loongarch64/lagoon_cpp.h
 
         common/spin_lock_loongarch64.cpp
     )

src/dynarmic/src/dynarmic/backend/loongarch64/a32_address_space.cpp

@@ -0,0 +1,123 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include "dynarmic/backend/loongarch64/a32_address_space.h"
+
+#include "common/assert.h"
+
+#include "dynarmic/backend/loongarch64/a32_jitstate.h"
+#include "dynarmic/backend/loongarch64/abi.h"
+#include "dynarmic/backend/loongarch64/emit_loongarch64.h"
+#include "dynarmic/frontend/A32/a32_location_descriptor.h"
+#include "dynarmic/frontend/A32/translate/a32_translate.h"
+#include "dynarmic/ir/opt_passes.h"
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+A32AddressSpace::A32AddressSpace(const A32::UserConfig& conf)
+        : conf(conf)
+        , cb(conf.code_cache_size) {
+    EmitPrelude();
+}
+
+void A32AddressSpace::GenerateIR(IR::Block& ir_block, IR::LocationDescriptor descriptor) const {
+    A32::Translate(ir_block, A32::LocationDescriptor{descriptor}, conf.callbacks, {conf.arch_version, conf.define_unpredictable_behaviour, conf.hook_hint_instructions});
+    Optimization::Optimize(ir_block, conf, {.sha256 = true});
+}
+
+CodePtr A32AddressSpace::Get(IR::LocationDescriptor descriptor) {
+    if (const auto iter = block_entries.find(descriptor.Value()); iter != block_entries.end()) {
+        return iter->second;
+    }
+    return nullptr;
+}
+
+CodePtr A32AddressSpace::GetOrEmit(IR::LocationDescriptor descriptor) {
+    if (CodePtr block_entry = Get(descriptor)) {
+        return block_entry;
+    }
+
+    IR::Block ir_block{descriptor};
+    GenerateIR(ir_block, descriptor);
+    const EmittedBlockInfo block_info = Emit(std::move(ir_block));
+
+    block_infos.insert_or_assign(descriptor.Value(), block_info);
+    block_entries.insert_or_assign(descriptor.Value(), block_info.entry_point);
+    return block_info.entry_point;
+}
+
+void A32AddressSpace::ClearCache() {
+    block_entries.clear();
+    block_infos.clear();
+    SetCursorPtr(prelude_info.end_of_prelude);
+}
+
+void A32AddressSpace::EmitPrelude() {
+    prelude_info.run_code = GetCursorPtr<PreludeInfo::RunCodeFuncType>();
+
+    // Save all GPRs except sp (r3) and tp (r2)
+    la_addi_d(&cb.as, LA_SP, LA_SP, -64 * 8);
+    for (u32 i = 1; i < 32; i++) {
+        if (i == LA_SP || i == LA_TP)
+            continue;
+        la_st_d(&cb.as, static_cast<la_gpr_t>(i), LA_SP, static_cast<int32_t>(i * 8));
+    }
+
+    // Set up reserved registers and jump to block entry
+    la_move(&cb.as, Xstate, LA_A1);  // Xstate = state ptr
+    la_move(&cb.as, Xhalt, LA_A2);   // Xhalt  = halt reason ptr
+    la_jr(&cb.as, LA_A0);            // jump to block_entry
+
+    prelude_info.return_from_run_code = GetCursorPtr<CodePtr>();
+
+    // Restore all GPRs except sp and tp
+    for (u32 i = 1; i < 32; i++) {
+        if (i == LA_SP || i == LA_TP)
+            continue;
+        la_ld_d(&cb.as, static_cast<la_gpr_t>(i), LA_SP, static_cast<int32_t>(i * 8));
+    }
+    la_addi_d(&cb.as, LA_SP, LA_SP, 64 * 8);
+    la_ret(&cb.as);
+
+    prelude_info.end_of_prelude = GetCursorPtr<CodePtr>();
+}
+
+void A32AddressSpace::SetCursorPtr(CodePtr ptr) {
+    cb.as.cursor = reinterpret_cast<uint8_t*>(ptr);
+}
+
+size_t A32AddressSpace::GetRemainingSize() {
+    return la_get_remaining_buffer_size(&cb.as);
+}
+
+EmittedBlockInfo A32AddressSpace::Emit(IR::Block block) {
+    if (GetRemainingSize() < 1024 * 1024) {
+        ClearCache();
+    }
+
+    EmitConfig emit_conf{};
+    EmittedBlockInfo block_info = EmitLoongArch64(cb.as, std::move(block), emit_conf);
+    Link(block_info);
+
+    return block_info;
+}
+
+void A32AddressSpace::Link(EmittedBlockInfo& block_info) {
+    for (const auto& reloc : block_info.relocations) {
+        uint8_t* patch_location = reinterpret_cast<uint8_t*>(block_info.entry_point) + reloc.code_offset;
+
+        switch (reloc.target) {
+        case LinkTarget::ReturnFromRunCode: {
+            std::ptrdiff_t off = reinterpret_cast<uint8_t*>(prelude_info.return_from_run_code) - patch_location;
+            lagoon_assembler_t patch_as;
+            la_init_assembler(&patch_as, patch_location, 4);
+            la_b(&patch_as, static_cast<int32_t>(off));
+            break;
+        }
+        default:
+            ASSERT(false && "Invalid relocation target");
+        }
+    }
+}
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/a32_address_space.h

@@ -0,0 +1,68 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include "dynarmic/backend/loongarch64/lagoon_cpp.h"
+#include <ankerl/unordered_dense.h>
+
+#include "dynarmic/backend/loongarch64/code_block.h"
+#include "dynarmic/backend/loongarch64/emit_loongarch64.h"
+#include "dynarmic/interface/A32/config.h"
+#include "dynarmic/interface/halt_reason.h"
+#include "dynarmic/ir/basic_block.h"
+#include "dynarmic/ir/location_descriptor.h"
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+struct A32JitState;
+
+class A32AddressSpace final {
+public:
+    explicit A32AddressSpace(const A32::UserConfig& conf);
+
+    void GenerateIR(IR::Block& ir_block, IR::LocationDescriptor descriptor) const;
+    CodePtr Get(IR::LocationDescriptor descriptor);
+    CodePtr GetOrEmit(IR::LocationDescriptor descriptor);
+
+    void ClearCache();
+
+private:
+    void EmitPrelude();
+
+    template<typename T>
+    T GetCursorPtr() {
+        return reinterpret_cast<T>(cb.as.cursor);
+    }
+
+    template<typename T>
+    T GetCursorPtr() const {
+        return reinterpret_cast<T>(cb.as.cursor);
+    }
+
+    template<typename T>
+    T GetMemPtr() const {
+        return cb.ptr<T>();
+    }
+
+    void SetCursorPtr(CodePtr ptr);
+    size_t GetRemainingSize();
+    EmittedBlockInfo Emit(IR::Block block);
+    void Link(EmittedBlockInfo& block_info);
+
+    const A32::UserConfig conf;
+    CodeBlock cb;
+
+    ankerl::unordered_dense::map<u64, CodePtr> block_entries;
+    ankerl::unordered_dense::map<u64, EmittedBlockInfo> block_infos;
+
+public:
+    struct PreludeInfo {
+        CodePtr end_of_prelude;
+        using RunCodeFuncType = HaltReason (*)(CodePtr entry_point, A32JitState* context, volatile u32* halt_reason);
+        RunCodeFuncType run_code;
+        CodePtr return_from_run_code;
+    } prelude_info;
+};
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/a32_interface.cpp

@@ -1,101 +1,134 @@
 // SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
-#include <array>
 #include <memory>
-#include <string>
+#include <mutex>
-#include <utility>
 
+#include <boost/icl/interval_set.hpp>
 #include "common/assert.h"
 #include "common/common_types.h"
+
+#include "dynarmic/backend/loongarch64/a32_address_space.h"
+#include "dynarmic/backend/loongarch64/a32_jitstate.h"
+#include "dynarmic/common/atomic.h"
+#include "dynarmic/frontend/A32/a32_location_descriptor.h"
 #include "dynarmic/interface/A32/a32.h"
 
 namespace Dynarmic::A32 {
 
+using namespace Backend::LoongArch64;
+
 struct Jit::Impl final {
-    explicit Impl(UserConfig conf_) : conf(std::move(conf_)) {}
+    Impl(Jit* jit_interface, A32::UserConfig conf)
+            : jit_interface(jit_interface)
+            , conf(conf)
+            , current_address_space(conf) {}
 
     HaltReason Run() {
-        UNIMPLEMENTED();
+        ASSERT(!jit_interface->is_executing);
-        return halt_reason;
+        jit_interface->is_executing = true;
+
+        const auto location_descriptor = current_state.GetLocationDescriptor();
+        const auto entry_point = current_address_space.GetOrEmit(location_descriptor);
+        current_address_space.prelude_info.run_code(entry_point, &current_state, &halt_reason);
+
+        HaltReason hr = static_cast<HaltReason>(Atomic::Exchange(&halt_reason, 0));
+        jit_interface->is_executing = false;
+        return hr;
     }
 
     HaltReason Step() {
-        UNIMPLEMENTED();
+        ASSERT(!jit_interface->is_executing);
-        return halt_reason | HaltReason::Step;
+        jit_interface->is_executing = true;
+
+        const auto location_descriptor = A32::LocationDescriptor{current_state.GetLocationDescriptor()}.SetSingleStepping(true);
+        const auto entry_point = current_address_space.GetOrEmit(location_descriptor);
+        current_address_space.prelude_info.run_code(entry_point, &current_state, &halt_reason);
+
+        HaltReason hr = static_cast<HaltReason>(Atomic::Exchange(&halt_reason, 0));
+        jit_interface->is_executing = false;
+        return hr;
     }
 
     void ClearCache() {
+        std::unique_lock lock{invalidation_mutex};
+        invalidate_entire_cache = true;
         HaltExecution(HaltReason::CacheInvalidation);
     }
 
-    void InvalidateCacheRange(u32, std::size_t) {
+    void InvalidateCacheRange(u32 start_address, size_t length) {
+        std::unique_lock lock{invalidation_mutex};
+        invalid_cache_ranges.add(boost::icl::discrete_interval<u32>::closed(start_address, static_cast<u32>(start_address + length - 1)));
         HaltExecution(HaltReason::CacheInvalidation);
     }
 
     void Reset() {
-        regs = {};
+        current_state = {};
-        ext_regs = {};
-        cpsr = 0;
-        fpscr = 0;
-        halt_reason = {};
     }
 
     void HaltExecution(HaltReason hr) {
-        halt_reason |= hr;
+        Atomic::Or(&halt_reason, static_cast<u32>(hr));
     }
 
     void ClearHalt(HaltReason hr) {
-        halt_reason &= ~hr;
+        Atomic::And(&halt_reason, ~static_cast<u32>(hr));
     }
 
     std::array<u32, 16>& Regs() {
-        return regs;
+        return current_state.regs;
     }
 
     const std::array<u32, 16>& Regs() const {
-        return regs;
+        return current_state.regs;
     }
 
     std::array<u32, 64>& ExtRegs() {
-        return ext_regs;
+        return current_state.ext_regs;
     }
 
     const std::array<u32, 64>& ExtRegs() const {
-        return ext_regs;
+        return current_state.ext_regs;
     }
 
     u32 Cpsr() const {
-        return cpsr;
+        return current_state.Cpsr();
     }
 
     void SetCpsr(u32 value) {
-        cpsr = value;
+        current_state.SetCpsr(value);
     }
 
     u32 Fpscr() const {
-        return fpscr;
+        return current_state.Fpscr();
     }
 
     void SetFpscr(u32 value) {
-        fpscr = value;
+        current_state.SetFpscr(value);
     }
 
-    void ClearExclusiveState() {}
+    void ClearExclusiveState() {
+        current_state.exclusive_state = false;
+    }
 
     std::string Disassemble() const {
         return {};
     }
 
-    UserConfig conf;
+private:
-    std::array<u32, 16> regs{};
+    Jit* jit_interface;
-    std::array<u32, 64> ext_regs{};
+    A32::UserConfig conf;
-    u32 cpsr = 0;
+    A32JitState current_state{};
-    u32 fpscr = 0;
+    A32AddressSpace current_address_space;
-    HaltReason halt_reason{};
+
+    volatile u32 halt_reason = 0;
+
+    std::mutex invalidation_mutex;
+    boost::icl::interval_set<u32> invalid_cache_ranges;
+    bool invalidate_entire_cache = false;
 };
 
-Jit::Jit(UserConfig conf) : impl(std::make_unique<Impl>(std::move(conf))) {}
+Jit::Jit(UserConfig conf)
+        : impl(std::make_unique<Impl>(this, conf)) {}
 
 Jit::~Jit() = default;
 
@@ -111,7 +144,7 @@ void Jit::ClearCache() {
     impl->ClearCache();
 }
 
-void Jit::InvalidateCacheRange(u32 start_address, std::size_t length) {
+void Jit::InvalidateCacheRange(u32 start_address, size_t length) {
     impl->InvalidateCacheRange(start_address, length);
 }
 

src/dynarmic/src/dynarmic/backend/loongarch64/a32_jitstate.cpp

@@ -0,0 +1,57 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include "dynarmic/backend/loongarch64/a32_jitstate.h"
+
+#include "dynarmic/mcl/bit.hpp"
+#include "common/common_types.h"
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+u32 A32JitState::Cpsr() const {
+    u32 cpsr = 0;
+    cpsr |= cpsr_nzcv;
+    cpsr |= cpsr_q;
+    cpsr |= mcl::bit::get_bit<31>(cpsr_ge) ? 1 << 19 : 0;
+    cpsr |= mcl::bit::get_bit<23>(cpsr_ge) ? 1 << 18 : 0;
+    cpsr |= mcl::bit::get_bit<15>(cpsr_ge) ? 1 << 17 : 0;
+    cpsr |= mcl::bit::get_bit<7>(cpsr_ge) ? 1 << 16 : 0;
+    cpsr |= mcl::bit::get_bit<1>(upper_location_descriptor) ? 1 << 9 : 0;
+    cpsr |= mcl::bit::get_bit<0>(upper_location_descriptor) ? 1 << 5 : 0;
+    cpsr |= static_cast<u32>(upper_location_descriptor & 0b11111100'00000000);
+    cpsr |= static_cast<u32>(upper_location_descriptor & 0b00000011'00000000) << 17;
+    cpsr |= cpsr_jaifm;
+    return cpsr;
+}
+
+void A32JitState::SetCpsr(u32 cpsr) {
+    cpsr_nzcv = cpsr & 0xF0000000;
+    cpsr_q = cpsr & (1 << 27);
+    cpsr_ge = 0;
+    cpsr_ge |= mcl::bit::get_bit<19>(cpsr) ? 0xFF000000 : 0;
+    cpsr_ge |= mcl::bit::get_bit<18>(cpsr) ? 0x00FF0000 : 0;
+    cpsr_ge |= mcl::bit::get_bit<17>(cpsr) ? 0x0000FF00 : 0;
+    cpsr_ge |= mcl::bit::get_bit<16>(cpsr) ? 0x000000FF : 0;
+
+    upper_location_descriptor &= 0xFFFF0000;
+    upper_location_descriptor |= mcl::bit::get_bit<9>(cpsr) ? 2 : 0;
+    upper_location_descriptor |= mcl::bit::get_bit<5>(cpsr) ? 1 : 0;
+    upper_location_descriptor |= (cpsr >> 0) & 0b11111100'00000000;
+    upper_location_descriptor |= (cpsr >> 17) & 0b00000011'00000000;
+    cpsr_jaifm = cpsr & 0x010001DF;
+}
+
+constexpr u32 FPCR_MASK = A32::LocationDescriptor::FPSCR_MODE_MASK;
+constexpr u32 FPSR_MASK = 0x0800009F;
+
+u32 A32JitState::Fpscr() const {
+    return (upper_location_descriptor & FPCR_MASK) | fpsr | fpsr_nzcv;
+}
+
+void A32JitState::SetFpscr(u32 fpscr) {
+    fpsr = fpscr & FPSR_MASK;
+    fpsr_nzcv = fpscr & 0xF0000000;
+    upper_location_descriptor = (upper_location_descriptor & 0x0000ffff) | (fpscr & FPCR_MASK);
+}
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/a32_jitstate.h

@@ -0,0 +1,43 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <array>
+
+#include "common/common_types.h"
+
+#include "dynarmic/frontend/A32/a32_location_descriptor.h"
+#include "dynarmic/ir/location_descriptor.h"
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+struct A32JitState {
+    u32 cpsr_nzcv = 0;
+    u32 cpsr_q = 0;
+    u32 cpsr_jaifm = 0;
+    u32 cpsr_ge = 0;
+
+    u32 fpsr = 0;
+    u32 fpsr_nzcv = 0;
+
+    std::array<u32, 16> regs{};
+
+    u32 upper_location_descriptor;
+
+    alignas(16) std::array<u32, 64> ext_regs{};
+
+    u32 exclusive_state = 0;
+
+    u32 Cpsr() const;
+    void SetCpsr(u32 cpsr);
+
+    u32 Fpscr() const;
+    void SetFpscr(u32 fpscr);
+
+    IR::LocationDescriptor GetLocationDescriptor() const {
+        return IR::LocationDescriptor{regs[15] | (static_cast<u64>(upper_location_descriptor) << 32)};
+    }
+};
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/abi.h

@@ -0,0 +1,39 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <initializer_list>
+
+#include "dynarmic/backend/loongarch64/lagoon_cpp.h"
+
+#include "common/common_types.h"
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+// Reserved registers used by the dynarmic JIT
+// Xstate (s0/r23): pointer to JIT CPU state, callee-saved
+// Xhalt  (s1/r24): halt reason pointer, callee-saved
+// Xscratch0 (t7/r19): scratch register (caller-saved)
+// Xscratch1 (t8/r20): scratch register (caller-saved)
+constexpr la_gpr_t Xstate    = LA_S0;   // r23
+constexpr la_gpr_t Xhalt     = LA_S1;   // r24
+constexpr la_gpr_t Xscratch0 = LA_T7;   // r19
+constexpr la_gpr_t Xscratch1 = LA_T8;   // r20
+
+// GPR allocation order: callee-saved (s2-s8) first, then temps, then args
+constexpr std::initializer_list<u32> GPR_ORDER{
+    25, 26, 27, 28, 29, 30, 31,   // s2-s8 (r25-r31)
+    12, 13, 14, 15, 16, 17, 18,   // t0-t6 (r12-r18)
+    4,  5,  6,  7,  8,  9, 10, 11 // a0-a7 (r4-r11)
+};
+
+// FPR allocation order: callee-saved (fs0-fs7) first, then ft, then fa
+constexpr std::initializer_list<u32> FPR_ORDER{
+    24, 25, 26, 27, 28, 29, 30, 31,   // fs0-fs7 (f24-f31)
+    8,  9, 10, 11, 12, 13, 14, 15,    // ft0-ft7 (f8-f15)
+    16, 17, 18, 19, 20, 21, 22, 23,   // ft8-ft15 (f16-f23)
+    0,  1,  2,  3,  4,  5,  6,  7    // fa0-fa7 (f0-f7)
+};
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/code_block.h

@@ -4,20 +4,44 @@
 #pragma once
 
 #include <cstdint>
+#include <new>
+#include <type_traits>
 
+#include <sys/mman.h>
+#include "dynarmic/backend/loongarch64/lagoon_cpp.h"
+
+#include "common/assert.h"
 #include "common/common_types.h"
 
 namespace Dynarmic::Backend::LoongArch64 {
 
 class CodeBlock {
 public:
+    explicit CodeBlock(std::size_t size) noexcept
+            : memsize(size) {
+        mem = static_cast<u8*>(mmap(nullptr, size, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_ANON | MAP_PRIVATE, -1, 0));
+        ASSERT(mem != MAP_FAILED);
+        la_init_assembler(&as, mem, size);
+    }
+
+    ~CodeBlock() noexcept {
+        if (mem == nullptr) {
+            return;
+        }
+        munmap(mem, memsize);
+    }
+
     template<typename T>
     T ptr() const noexcept {
+        static_assert(std::is_pointer_v<T> || std::is_same_v<T, std::uintptr_t> || std::is_same_v<T, std::intptr_t>);
         return reinterpret_cast<T>(mem);
     }
 
+    lagoon_assembler_t as{};
+
 private:
     u8* mem = nullptr;
+    size_t memsize = 0;
 };
 
 }  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/emit_context.h

@@ -0,0 +1,24 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include "dynarmic/backend/loongarch64/emit_loongarch64.h"
+#include "dynarmic/backend/loongarch64/reg_alloc.h"
+
+namespace Dynarmic::IR {
+class Block;
+}  // namespace Dynarmic::IR
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+struct EmitConfig;
+
+struct EmitContext {
+    IR::Block& block;
+    RegAlloc& reg_alloc;
+    const EmitConfig& emit_conf;
+    EmittedBlockInfo& ebi;
+};
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/emit_loongarch64.cpp

@@ -0,0 +1,113 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include "dynarmic/backend/loongarch64/emit_loongarch64.h"
+
+#include "dynarmic/backend/loongarch64/a32_jitstate.h"
+#include "dynarmic/backend/loongarch64/abi.h"
+#include "dynarmic/backend/loongarch64/emit_context.h"
+#include "dynarmic/backend/loongarch64/reg_alloc.h"
+#include "dynarmic/ir/basic_block.h"
+#include "dynarmic/ir/microinstruction.h"
+#include "dynarmic/ir/opcodes.h"
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+template<IR::Opcode op>
+void EmitIR(lagoon_assembler_t&, EmitContext&, IR::Inst*) {
+    ASSERT(false && "Unimplemented opcode");
+}
+
+template<>
+void EmitIR<IR::Opcode::Void>(lagoon_assembler_t&, EmitContext&, IR::Inst*) {}
+
+template<>
+void EmitIR<IR::Opcode::A32GetRegister>(lagoon_assembler_t&, EmitContext& ctx, IR::Inst* inst);
+
+template<>
+void EmitIR<IR::Opcode::A32SetRegister>(lagoon_assembler_t&, EmitContext& ctx, IR::Inst* inst);
+
+template<>
+void EmitIR<IR::Opcode::A32SetCpsrNZC>(lagoon_assembler_t&, EmitContext& ctx, IR::Inst* inst);
+
+template<>
+void EmitIR<IR::Opcode::LogicalShiftLeft32>(lagoon_assembler_t&, EmitContext& ctx, IR::Inst* inst);
+
+template<>
+void EmitIR<IR::Opcode::GetCarryFromOp>(lagoon_assembler_t&, EmitContext& ctx, IR::Inst* inst) {
+    ASSERT(ctx.reg_alloc.IsValueLive(inst));
+}
+
+template<>
+void EmitIR<IR::Opcode::GetNZFromOp>(lagoon_assembler_t& as, EmitContext& ctx, IR::Inst* inst) {
+    auto args = ctx.reg_alloc.GetArgumentInfo(inst);
+
+    auto Xvalue = ctx.reg_alloc.ReadX(args[0]);
+    auto Xnz = ctx.reg_alloc.WriteX(inst);
+    RegAlloc::Realize(Xvalue, Xnz);
+
+    // Z flag (bit 30): set if value == 0
+    la_sltui(&as, Xnz->index, Xvalue->index, 1);
+    la_slli_d(&as, Xnz->index, Xnz->index, 30);
+    // N flag (bit 31): set if value < 0 (signed)
+    la_slt(&as, Xscratch0, Xvalue->index, LA_ZERO);
+    la_slli_d(&as, Xscratch0, Xscratch0, 31);
+    la_or(&as, Xnz->index, Xnz->index, Xscratch0);
+}
+
+EmittedBlockInfo EmitLoongArch64(lagoon_assembler_t& as, IR::Block block, const EmitConfig& emit_conf) {
+    EmittedBlockInfo ebi;
+
+    RegAlloc reg_alloc{as, std::vector<u32>(GPR_ORDER), std::vector<u32>(FPR_ORDER)};
+    EmitContext ctx{block, reg_alloc, emit_conf, ebi};
+
+    ebi.entry_point = reinterpret_cast<CodePtr>(as.cursor);
+
+    for (auto iter = block.instructions.begin(); iter != block.instructions.end(); ++iter) {
+        IR::Inst* inst = &*iter;
+
+        switch (inst->GetOpcode()) {
+#define OPCODE(name, type, ...)                            \
+    case IR::Opcode::name:                                 \
+        EmitIR<IR::Opcode::name>(as, ctx, inst);           \
+        break;
+#define A32OPC(name, type, ...)                                 \
+    case IR::Opcode::A32##name:                                 \
+        EmitIR<IR::Opcode::A32##name>(as, ctx, inst);           \
+        break;
+#define A64OPC(name, type, ...)                                 \
+    case IR::Opcode::A64##name:                                 \
+        EmitIR<IR::Opcode::A64##name>(as, ctx, inst);           \
+        break;
+#include "dynarmic/ir/opcodes.inc"
+#undef OPCODE
+#undef A32OPC
+#undef A64OPC
+        default:
+            UNREACHABLE();
+        }
+
+        reg_alloc.UpdateAllUses();
+    }
+
+    reg_alloc.UpdateAllUses();
+    reg_alloc.AssertNoMoreUses();
+
+    // TODO: Emit Terminal
+    const auto term = block.GetTerminal();
+    const IR::Term::LinkBlock* link_block_term = boost::get<IR::Term::LinkBlock>(&term);
+    ASSERT(link_block_term);
+    la_load_immediate64(&as, Xscratch0, link_block_term->next.Value());
+    la_st_w(&as, Xscratch0, Xstate, static_cast<int32_t>(offsetof(A32JitState, regs) + sizeof(u32) * 15));
+
+    ebi.relocations.push_back(Relocation{
+        reinterpret_cast<CodePtr>(as.cursor) - ebi.entry_point,
+        LinkTarget::ReturnFromRunCode
+    });
+    la_nop(&as);
+
+    ebi.size = reinterpret_cast<CodePtr>(as.cursor) - ebi.entry_point;
+    return ebi;
+}
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/emit_loongarch64.h

@@ -0,0 +1,55 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <array>
+#include <cstddef>
+#include <vector>
+
+#include "dynarmic/backend/loongarch64/lagoon_cpp.h"
+
+#include "common/common_types.h"
+
+namespace Dynarmic::IR {
+class Block;
+class Inst;
+class LocationDescriptor;
+enum class Cond;
+enum class Opcode;
+}  // namespace Dynarmic::IR
+
+namespace Dynarmic::A32 {
+class Coprocessor;
+}  // namespace Dynarmic::A32
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+using CodePtr = std::byte*;
+
+enum class LinkTarget {
+    ReturnFromRunCode,
+};
+
+struct Relocation {
+    std::ptrdiff_t code_offset;
+    LinkTarget target;
+};
+
+struct EmittedBlockInfo {
+    std::vector<Relocation> relocations;
+    CodePtr entry_point;
+    size_t size;
+    size_t cycle_count;
+};
+
+struct EmitConfig {};
+
+struct EmitContext;
+
+template<IR::Opcode op>
+void EmitIR(lagoon_assembler_t& as, EmitContext& ctx, IR::Inst* inst);
+
+EmittedBlockInfo EmitLoongArch64(lagoon_assembler_t& as, IR::Block block, const EmitConfig& emit_conf);
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/emit_loongarch64_a32.cpp

@@ -0,0 +1,59 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include "dynarmic/backend/loongarch64/lagoon_cpp.h"
+
+#include "dynarmic/backend/loongarch64/a32_jitstate.h"
+#include "dynarmic/backend/loongarch64/abi.h"
+#include "dynarmic/backend/loongarch64/emit_context.h"
+#include "dynarmic/backend/loongarch64/emit_loongarch64.h"
+#include "dynarmic/backend/loongarch64/reg_alloc.h"
+#include "dynarmic/ir/basic_block.h"
+#include "dynarmic/ir/microinstruction.h"
+#include "dynarmic/ir/opcodes.h"
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+template<>
+void EmitIR<IR::Opcode::A32GetRegister>(lagoon_assembler_t& as, EmitContext& ctx, IR::Inst* inst) {
+    const A32::Reg reg = inst->GetArg(0).GetA32RegRef();
+
+    auto Xresult = ctx.reg_alloc.WriteX(inst);
+    RegAlloc::Realize(Xresult);
+
+    la_ld_wu(&as, Xresult->index, Xstate,
+             static_cast<int32_t>(offsetof(A32JitState, regs) + sizeof(u32) * static_cast<size_t>(reg)));
+}
+
+template<>
+void EmitIR<IR::Opcode::A32SetRegister>(lagoon_assembler_t& as, EmitContext& ctx, IR::Inst* inst) {
+    const A32::Reg reg = inst->GetArg(0).GetA32RegRef();
+
+    auto args = ctx.reg_alloc.GetArgumentInfo(inst);
+    auto Xvalue = ctx.reg_alloc.ReadX(args[1]);
+    RegAlloc::Realize(Xvalue);
+
+    la_st_w(&as, Xvalue->index, Xstate,
+            static_cast<int32_t>(offsetof(A32JitState, regs) + sizeof(u32) * static_cast<size_t>(reg)));
+}
+
+template<>
+void EmitIR<IR::Opcode::A32SetCpsrNZC>(lagoon_assembler_t& as, EmitContext& ctx, IR::Inst* inst) {
+    auto args = ctx.reg_alloc.GetArgumentInfo(inst);
+
+    ASSERT(!args[0].IsImmediate() && !args[1].IsImmediate());
+
+    auto Xnz = ctx.reg_alloc.ReadX(args[0]);
+    auto Xc = ctx.reg_alloc.ReadX(args[1]);
+    RegAlloc::Realize(Xnz, Xc);
+
+    la_ld_wu(&as, Xscratch0, Xstate, static_cast<int32_t>(offsetof(A32JitState, cpsr_nzcv)));
+    la_load_immediate64(&as, Xscratch1, 0x10000000);
+    la_and(&as, Xscratch0, Xscratch0, Xscratch1);
+    la_or(&as, Xscratch0, Xscratch0, Xnz->index);
+    la_slli_w(&as, Xscratch1, Xc->index, 29);
+    la_or(&as, Xscratch0, Xscratch0, Xscratch1);
+    la_st_w(&as, Xscratch0, Xstate, static_cast<int32_t>(offsetof(A32JitState, cpsr_nzcv)));
+}
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/emit_loongarch64_data_processing.cpp

@@ -0,0 +1,52 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include "dynarmic/backend/loongarch64/lagoon_cpp.h"
+
+#include "dynarmic/backend/loongarch64/abi.h"
+#include "dynarmic/backend/loongarch64/emit_context.h"
+#include "dynarmic/backend/loongarch64/emit_loongarch64.h"
+#include "dynarmic/backend/loongarch64/reg_alloc.h"
+#include "dynarmic/ir/basic_block.h"
+#include "dynarmic/ir/microinstruction.h"
+#include "dynarmic/ir/opcodes.h"
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+template<>
+void EmitIR<IR::Opcode::LogicalShiftLeft32>(lagoon_assembler_t& as, EmitContext& ctx, IR::Inst* inst) {
+    const auto carry_inst = inst->GetAssociatedPseudoOperation(IR::Opcode::GetCarryFromOp);
+
+    auto args = ctx.reg_alloc.GetArgumentInfo(inst);
+    auto& operand_arg = args[0];
+    auto& shift_arg = args[1];
+    auto& carry_arg = args[2];
+
+    ASSERT(carry_inst != nullptr);
+    ASSERT(shift_arg.IsImmediate());
+
+    auto Xresult = ctx.reg_alloc.WriteX(inst);
+    auto Xcarry_out = ctx.reg_alloc.WriteX(carry_inst);
+    auto Xoperand = ctx.reg_alloc.ReadX(operand_arg);
+    auto Xcarry_in = ctx.reg_alloc.ReadX(carry_arg);
+    RegAlloc::Realize(Xresult, Xcarry_out, Xoperand, Xcarry_in);
+
+    const u8 shift = shift_arg.GetImmediateU8();
+
+    if (shift == 0) {
+        la_addi_w(&as, Xresult->index, Xoperand->index, 0);
+        la_addi_w(&as, Xcarry_out->index, Xcarry_in->index, 0);
+    } else if (shift < 32) {
+        la_srli_w(&as, Xcarry_out->index, Xoperand->index, 32 - shift);
+        la_andi(&as, Xcarry_out->index, Xcarry_out->index, 1);
+        la_slli_w(&as, Xresult->index, Xoperand->index, shift);
+    } else if (shift > 32) {
+        la_move(&as, Xresult->index, LA_ZERO);
+        la_move(&as, Xcarry_out->index, LA_ZERO);
+    } else {
+        la_andi(&as, Xcarry_out->index, Xoperand->index, 1);
+        la_move(&as, Xresult->index, LA_ZERO);
+    }
+}
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/lagoon_cpp.h

@@ -0,0 +1,8 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+extern "C" {
+#include <lagoon.h>
+}

src/dynarmic/src/dynarmic/backend/loongarch64/reg_alloc.cpp

@@ -0,0 +1,364 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include "dynarmic/backend/loongarch64/reg_alloc.h"
+
+#include <algorithm>
+#include <array>
+#include <limits>
+
+#include "dynarmic/backend/loongarch64/lagoon_cpp.h"
+
+#include "common/assert.h"
+#include "common/common_types.h"
+
+#include "dynarmic/common/always_false.h"
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+constexpr size_t spill_offset = offsetof(StackLayout, spill);
+constexpr size_t spill_slot_size = sizeof(decltype(StackLayout::spill)::value_type);
+
+static bool IsValuelessType(IR::Type type) {
+    switch (type) {
+    case IR::Type::Table:
+        return true;
+    default:
+        return false;
+    }
+}
+
+IR::Type Argument::GetType() const {
+    return value.GetType();
+}
+
+bool Argument::IsImmediate() const {
+    return value.IsImmediate();
+}
+
+bool Argument::GetImmediateU1() const {
+    return value.GetU1();
+}
+
+u8 Argument::GetImmediateU8() const {
+    const u64 imm = value.GetImmediateAsU64();
+    ASSERT(imm < 0x100);
+    return u8(imm);
+}
+
+u16 Argument::GetImmediateU16() const {
+    const u64 imm = value.GetImmediateAsU64();
+    ASSERT(imm < 0x10000);
+    return u16(imm);
+}
+
+u32 Argument::GetImmediateU32() const {
+    const u64 imm = value.GetImmediateAsU64();
+    ASSERT(imm < 0x100000000);
+    return u32(imm);
+}
+
+u64 Argument::GetImmediateU64() const {
+    return value.GetImmediateAsU64();
+}
+
+IR::Cond Argument::GetImmediateCond() const {
+    ASSERT(IsImmediate() && GetType() == IR::Type::Cond);
+    return value.GetCond();
+}
+
+IR::AccType Argument::GetImmediateAccType() const {
+    ASSERT(IsImmediate() && GetType() == IR::Type::AccType);
+    return value.GetAccType();
+}
+
+bool HostLocInfo::Contains(const IR::Inst* value) const {
+    return std::find(values.begin(), values.end(), value) != values.end();
+}
+
+void HostLocInfo::SetupScratchLocation() {
+    ASSERT(IsCompletelyEmpty());
+    locked = 1;
+    realized = true;
+}
+
+bool HostLocInfo::IsCompletelyEmpty() const {
+    return values.empty() && !locked && !accumulated_uses && !expected_uses && !uses_this_inst && !realized;
+}
+
+void HostLocInfo::UpdateUses() {
+    accumulated_uses += uses_this_inst;
+    uses_this_inst = 0;
+
+    if (accumulated_uses == expected_uses) {
+        values.clear();
+        accumulated_uses = 0;
+        expected_uses = 0;
+    }
+}
+
+RegAlloc::ArgumentInfo RegAlloc::GetArgumentInfo(IR::Inst* inst) {
+    ArgumentInfo ret = {Argument{}, Argument{}, Argument{}, Argument{}};
+    for (size_t i = 0; i < inst->NumArgs(); i++) {
+        const IR::Value arg = inst->GetArg(i);
+        ret[i].value = arg;
+        if (!arg.IsImmediate() && !IsValuelessType(arg.GetType())) {
+            ASSERT(ValueLocation(arg.GetInst()) && "argument must already been defined");
+            ValueInfo(arg.GetInst()).uses_this_inst++;
+        }
+    }
+    return ret;
+}
+
+bool RegAlloc::IsValueLive(IR::Inst* inst) const {
+    return !!ValueLocation(inst);
+}
+
+void RegAlloc::UpdateAllUses() {
+    for (auto& info : hostloc_info) {
+        info.UpdateUses();
+    }
+}
+
+void RegAlloc::DefineAsExisting(IR::Inst* inst, Argument& arg) {
+    ASSERT(!ValueLocation(inst));
+
+    if (arg.value.IsImmediate()) {
+        inst->ReplaceUsesWith(arg.value);
+        return;
+    }
+
+    auto& info = ValueInfo(arg.value.GetInst());
+    info.values.emplace_back(inst);
+    info.expected_uses += inst->UseCount();
+}
+
+void RegAlloc::AssertNoMoreUses() const {
+    // TODO: Re-enable this assert once all register allocation issues are fixed
+    // const auto is_empty = [](const auto& i) { return i.IsCompletelyEmpty(); };
+    // ASSERT(std::all_of(hostloc_info.begin(), hostloc_info.end(), is_empty));
+}
+
+template<HostLoc::Kind kind>
+u32 RegAlloc::GenerateImmediate(const IR::Value& value) {
+    if constexpr (kind == HostLoc::Kind::Gpr) {
+        const u32 new_location_index = AllocateRegister(gpr_order, GprOffset);
+        SpillGpr(new_location_index);
+        hostloc_info[GprOffset + new_location_index].SetupScratchLocation();
+
+        la_load_immediate64(&as, static_cast<la_gpr_t>(new_location_index),
+                            static_cast<int64_t>(value.GetImmediateAsU64()));
+
+        return new_location_index;
+    } else if constexpr (kind == HostLoc::Kind::Fpr) {
+        ASSERT(false && "Unimplemented instruction");
+    } else {
+        UNREACHABLE();
+    }
+    return 0;
+}
+
+template<HostLoc::Kind required_kind>
+u32 RegAlloc::RealizeReadImpl(const IR::Value& value) {
+    if (value.IsImmediate()) {
+        return GenerateImmediate<required_kind>(value);
+    }
+
+    const auto current_location = ValueLocation(value.GetInst());
+    ASSERT(current_location);
+
+    if (current_location->kind == required_kind) {
+        ValueInfo(*current_location).realized = true;
+        return current_location->index;
+    }
+
+    ASSERT(!ValueInfo(*current_location).realized);
+    ASSERT(!ValueInfo(*current_location).locked);
+
+    if constexpr (required_kind == HostLoc::Kind::Gpr) {
+        const u32 new_location_index = AllocateRegister(gpr_order, GprOffset);
+        SpillGpr(new_location_index);
+
+        switch (current_location->kind) {
+        case HostLoc::Kind::Gpr:
+            UNREACHABLE();
+        case HostLoc::Kind::Fpr:
+            la_movfr2gr_d(&as, static_cast<la_gpr_t>(new_location_index),
+                          static_cast<la_fpr_t>(current_location->index));
+            break;
+        case HostLoc::Kind::Spill:
+            la_ld_d(&as, static_cast<la_gpr_t>(new_location_index), LA_SP,
+                    static_cast<int32_t>(spill_offset + current_location->index * spill_slot_size));
+            break;
+        }
+
+        hostloc_info[GprOffset + new_location_index] = std::exchange(ValueInfo(*current_location), {});
+        hostloc_info[GprOffset + new_location_index].realized = true;
+        return new_location_index;
+    } else if constexpr (required_kind == HostLoc::Kind::Fpr) {
+        const u32 new_location_index = AllocateRegister(fpr_order, FprOffset);
+        SpillFpr(new_location_index);
+
+        switch (current_location->kind) {
+        case HostLoc::Kind::Gpr:
+            la_movgr2fr_d(&as, static_cast<la_fpr_t>(new_location_index),
+                          static_cast<la_gpr_t>(current_location->index));
+            break;
+        case HostLoc::Kind::Fpr:
+            UNREACHABLE();
+        case HostLoc::Kind::Spill:
+            la_fld_d(&as, static_cast<la_fpr_t>(new_location_index), LA_SP,
+                     static_cast<int32_t>(spill_offset + current_location->index * spill_slot_size));
+            break;
+        }
+
+        hostloc_info[FprOffset + new_location_index] = std::exchange(ValueInfo(*current_location), {});
+        hostloc_info[FprOffset + new_location_index].realized = true;
+        return new_location_index;
+    } else {
+        UNREACHABLE();
+    }
+}
+
+u32 RegAlloc::RealizeWriteImpl(const IR::Inst* value, HostLoc::Kind required_kind) {
+    if (value == nullptr) {
+        // Scratch register allocation
+        if (required_kind == HostLoc::Kind::Gpr) {
+            const u32 idx = AllocateRegister(gpr_order, GprOffset);
+            SpillGpr(idx);
+            hostloc_info[GprOffset + idx].SetupScratchLocation();
+            return idx;
+        } else if (required_kind == HostLoc::Kind::Fpr) {
+            const u32 idx = AllocateRegister(fpr_order, FprOffset);
+            SpillFpr(idx);
+            hostloc_info[FprOffset + idx].SetupScratchLocation();
+            return idx;
+        }
+    }
+
+    ASSERT(!ValueLocation(value));
+
+    const auto setup_location = [&](HostLocInfo& info) {
+        info = {};
+        info.values.emplace_back(value);
+        info.locked = true;
+        info.realized = true;
+        info.expected_uses = value->UseCount();
+    };
+
+    if (required_kind == HostLoc::Kind::Gpr) {
+        const u32 new_location_index = AllocateRegister(gpr_order, GprOffset);
+        SpillGpr(new_location_index);
+        setup_location(hostloc_info[GprOffset + new_location_index]);
+        return new_location_index;
+    } else if (required_kind == HostLoc::Kind::Fpr) {
+        const u32 new_location_index = AllocateRegister(fpr_order, FprOffset);
+        SpillFpr(new_location_index);
+        setup_location(hostloc_info[FprOffset + new_location_index]);
+        return new_location_index;
+    } else {
+        UNREACHABLE();
+    }
+}
+
+template u32 RegAlloc::RealizeReadImpl<HostLoc::Kind::Gpr>(const IR::Value& value);
+template u32 RegAlloc::RealizeReadImpl<HostLoc::Kind::Fpr>(const IR::Value& value);
+
+u32 RegAlloc::AllocateRegister(const std::vector<u32>& order, size_t base_offset) {
+    const auto empty = std::find_if(order.begin(), order.end(), [&](u32 i) {
+        auto& info = hostloc_info[base_offset + i];
+        return info.values.empty() && !info.locked;
+    });
+    if (empty != order.end()) {
+        return *empty;
+    }
+
+    std::vector<u32> candidates;
+    std::copy_if(order.begin(), order.end(), std::back_inserter(candidates), [&](u32 i) {
+        return !hostloc_info[base_offset + i].locked;
+    });
+    ASSERT(!candidates.empty());
+
+    u32 best = candidates[0];
+    size_t min_lru = hostloc_info[base_offset + best].lru_counter;
+    for (size_t i = 1; i < candidates.size(); ++i) {
+        auto& info = hostloc_info[base_offset + candidates[i]];
+        if (info.lru_counter < min_lru) {
+            min_lru = info.lru_counter;
+            best = candidates[i];
+        }
+    }
+    hostloc_info[base_offset + best].lru_counter++;
+    return best;
+}
+
+void RegAlloc::SpillGpr(u32 index) {
+    auto& gpr_info = hostloc_info[GprOffset + index];
+    ASSERT(!gpr_info.locked && !gpr_info.realized);
+    if (gpr_info.values.empty()) {
+        return;
+    }
+    const u32 new_location_index = FindFreeSpill();
+    la_st_d(&as, static_cast<la_gpr_t>(index), LA_SP,
+            static_cast<int32_t>(spill_offset + new_location_index * spill_slot_size));
+    hostloc_info[SpillOffset + new_location_index] = std::exchange(gpr_info, {});
+}
+
+void RegAlloc::SpillFpr(u32 index) {
+    auto& fpr_info = hostloc_info[FprOffset + index];
+    ASSERT(!fpr_info.locked && !fpr_info.realized);
+    if (fpr_info.values.empty()) {
+        return;
+    }
+    const u32 new_location_index = FindFreeSpill();
+    la_fst_d(&as, static_cast<la_fpr_t>(index), LA_SP,
+             static_cast<int32_t>(spill_offset + new_location_index * spill_slot_size));
+    hostloc_info[SpillOffset + new_location_index] = std::exchange(fpr_info, {});
+}
+
+u32 RegAlloc::FindFreeSpill() const {
+    for (size_t i = 0; i < SpillCount; ++i) {
+        if (hostloc_info[SpillOffset + i].values.empty()) {
+            return static_cast<u32>(i);
+        }
+    }
+    UNREACHABLE();
+}
+
+std::optional<HostLoc> RegAlloc::ValueLocation(const IR::Inst* value) const {
+    for (size_t i = 0; i < hostloc_info.size(); ++i) {
+        if (hostloc_info[i].Contains(value)) {
+            if (i < GprCount) {
+                return HostLoc{HostLoc::Kind::Gpr, static_cast<u32>(i)};
+            } else if (i < GprCount + FprCount) {
+                return HostLoc{HostLoc::Kind::Fpr, static_cast<u32>(i - GprCount)};
+            } else {
+                return HostLoc{HostLoc::Kind::Spill, static_cast<u32>(i - GprCount - FprCount)};
+            }
+        }
+    }
+    return std::nullopt;
+}
+
+HostLocInfo& RegAlloc::ValueInfo(HostLoc host_loc) {
+    switch (host_loc.kind) {
+    case HostLoc::Kind::Gpr:
+        return hostloc_info[GprOffset + host_loc.index];
+    case HostLoc::Kind::Fpr:
+        return hostloc_info[FprOffset + host_loc.index];
+    case HostLoc::Kind::Spill:
+        return hostloc_info[SpillOffset + host_loc.index];
+    }
+    UNREACHABLE();
+}
+
+HostLocInfo& RegAlloc::ValueInfo(const IR::Inst* value) {
+    for (auto& info : hostloc_info) {
+        if (info.Contains(value)) {
+            return info;
+        }
+    }
+    UNREACHABLE();
+}
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/reg_alloc.h

@@ -0,0 +1,232 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <array>
+#include <optional>
+#include <utility>
+#include <vector>
+
+#include "dynarmic/backend/loongarch64/lagoon_cpp.h"
+#include <ankerl/unordered_dense.h>
+
+#include "common/assert.h"
+#include "common/common_types.h"
+#include "dynarmic/mcl/is_instance_of_template.hpp"
+
+#include "dynarmic/backend/loongarch64/stack_layout.h"
+#include "dynarmic/ir/cond.h"
+#include "dynarmic/ir/microinstruction.h"
+#include "dynarmic/ir/value.h"
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+class RegAlloc;
+
+// Wrapper types for LoongArch GPR/FPR (replacing biscuit's register types)
+struct GPR {
+    la_gpr_t index = LA_ZERO;
+    GPR() = default;
+    explicit GPR(u32 i) : index{static_cast<la_gpr_t>(i)} {}
+    uint32_t Index() const { return static_cast<uint32_t>(index); }
+};
+
+struct FPR {
+    la_fpr_t index = LA_F0;
+    FPR() = default;
+    explicit FPR(u32 i) : index{static_cast<la_fpr_t>(i)} {}
+    uint32_t Index() const { return static_cast<uint32_t>(index); }
+};
+
+struct VPR {
+    la_vpr_t index;
+    VPR() = default;
+    explicit VPR(u32 i) : index{static_cast<la_vpr_t>(i)} {}
+    uint32_t Index() const { return static_cast<uint32_t>(index); }
+};
+
+struct HostLoc {
+    enum class Kind {
+        Gpr,
+        Fpr,
+        Spill,
+    } kind;
+    u32 index;
+};
+
+struct Argument {
+public:
+    using copyable_reference = std::reference_wrapper<Argument>;
+
+    IR::Type GetType() const;
+    bool IsImmediate() const;
+
+    bool GetImmediateU1() const;
+    u8 GetImmediateU8() const;
+    u16 GetImmediateU16() const;
+    u32 GetImmediateU32() const;
+    u64 GetImmediateU64() const;
+    IR::Cond GetImmediateCond() const;
+    IR::AccType GetImmediateAccType() const;
+
+private:
+    friend class RegAlloc;
+    explicit Argument() {}
+
+    IR::Value value;
+    bool allocated = false;
+};
+
+template<typename T>
+struct RAReg {
+public:
+    static constexpr HostLoc::Kind kind = std::is_same_v<T, FPR> || std::is_same_v<T, VPR>
+                                            ? HostLoc::Kind::Fpr
+                                            : HostLoc::Kind::Gpr;
+
+    operator T() const { return *reg; }
+
+    T operator*() const { return *reg; }
+
+    const T* operator->() const { return &*reg; }
+
+    ~RAReg();
+
+private:
+    friend class RegAlloc;
+    explicit RAReg(RegAlloc& reg_alloc, bool write, const IR::Value& value);
+
+    void Realize();
+
+    RegAlloc& reg_alloc;
+    bool write;
+    const IR::Value value;
+    std::optional<T> reg;
+};
+
+struct HostLocInfo final {
+    std::vector<const IR::Inst*> values;
+    size_t locked = 0;
+    size_t uses_this_inst = 0;
+    size_t accumulated_uses = 0;
+    size_t expected_uses = 0;
+    bool realized = false;
+    size_t lru_counter = 0;
+
+    bool Contains(const IR::Inst*) const;
+    void SetupScratchLocation();
+    bool IsCompletelyEmpty() const;
+    void UpdateUses();
+};
+
+class RegAlloc {
+public:
+    using ArgumentInfo = std::array<Argument, IR::max_arg_count>;
+
+    explicit RegAlloc(lagoon_assembler_t& as, std::vector<u32> gpr_order, std::vector<u32> fpr_order)
+            : as{as}, gpr_order{std::move(gpr_order)}, fpr_order{std::move(fpr_order)} {}
+
+    ArgumentInfo GetArgumentInfo(IR::Inst* inst);
+    bool IsValueLive(IR::Inst* inst) const;
+
+    auto ReadX(Argument& arg) { return RAReg<GPR>{*this, false, arg.value}; }
+    auto ReadD(Argument& arg) { return RAReg<FPR>{*this, false, arg.value}; }
+    auto ReadV(Argument& arg) { return RAReg<VPR>{*this, false, arg.value}; }
+
+    auto WriteX(IR::Inst* inst) { return RAReg<GPR>{*this, true, inst ? IR::Value{inst} : IR::Value{}}; }
+    auto WriteD(IR::Inst* inst) { return RAReg<FPR>{*this, true, inst ? IR::Value{inst} : IR::Value{}}; }
+    auto WriteV(IR::Inst* inst) { return RAReg<VPR>{*this, true, inst ? IR::Value{inst} : IR::Value{}}; }
+
+    auto ScratchGpr() { return RAReg<GPR>{*this, true, IR::Value{}}; }
+    auto ScratchFpr() { return RAReg<FPR>{*this, true, IR::Value{}}; }
+    auto ScratchVec() { return RAReg<VPR>{*this, true, IR::Value{}}; }
+
+    void DefineAsExisting(IR::Inst* inst, Argument& arg);
+
+    template<typename... Ts>
+    static void Realize(Ts&... rs) {
+        static_assert((mcl::is_instance_of_template<RAReg, Ts>() && ...));
+        (rs.Realize(), ...);
+    }
+
+    void UpdateAllUses();
+    void AssertNoMoreUses() const;
+
+private:
+    template<typename>
+    friend struct RAReg;
+
+    template<HostLoc::Kind kind>
+    u32 GenerateImmediate(const IR::Value& value);
+    template<HostLoc::Kind kind>
+    u32 RealizeReadImpl(const IR::Value& value);
+    u32 RealizeWriteImpl(const IR::Inst* value, HostLoc::Kind required_kind);
+
+    u32 AllocateRegister(const std::vector<u32>& order, size_t base_offset);
+    void SpillGpr(u32 index);
+    void SpillFpr(u32 index);
+    u32 FindFreeSpill() const;
+
+    std::optional<HostLoc> ValueLocation(const IR::Inst* value) const;
+    HostLocInfo& ValueInfo(HostLoc host_loc);
+    HostLocInfo& ValueInfo(const IR::Inst* value);
+
+    lagoon_assembler_t& as;
+    std::vector<u32> gpr_order;
+    std::vector<u32> fpr_order;
+
+    static constexpr size_t GprCount = 32;
+    static constexpr size_t FprCount = 32;
+    static constexpr size_t GprOffset = 0;
+    static constexpr size_t FprOffset = GprCount;
+    static constexpr size_t SpillOffset = GprCount + FprCount;
+
+    std::array<HostLocInfo, GprCount + FprCount + SpillCount> hostloc_info;
+};
+
+template<typename T>
+RAReg<T>::RAReg(RegAlloc& reg_alloc, bool write, const IR::Value& value)
+        : reg_alloc{reg_alloc}, write{write}, value{value} {
+    if (!write && !value.IsImmediate()) {
+        reg_alloc.ValueInfo(value.GetInst()).locked++;
+    }
+}
+
+template<typename T>
+RAReg<T>::~RAReg() {
+    if (value.IsEmpty()) {
+        if (reg) {
+            HostLocInfo& info = reg_alloc.ValueInfo(HostLoc{kind, reg->Index()});
+            info.locked--;
+            info.realized = false;
+        }
+        return;
+    }
+
+    if (value.IsImmediate()) {
+        if (reg) {
+            // Immediate was materialized into a scratch register
+            HostLocInfo& info = reg_alloc.ValueInfo(HostLoc{kind, reg->Index()});
+            info.locked--;
+            info.realized = false;
+        }
+    } else if (!value.IsEmpty()) {
+        HostLocInfo& info = reg_alloc.ValueInfo(value.GetInst());
+        info.locked--;
+        if (reg) {
+            info.realized = false;
+        }
+    }
+}
+
+template<typename T>
+void RAReg<T>::Realize() {
+    if (write && value.IsEmpty()) {
+        reg = T{reg_alloc.RealizeWriteImpl(nullptr, kind)};
+    } else {
+        reg = T{write ? reg_alloc.RealizeWriteImpl(value.GetInst(), kind) : reg_alloc.RealizeReadImpl<kind>(value)};
+    }
+}
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/backend/loongarch64/stack_layout.h

@@ -0,0 +1,28 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <array>
+
+#include "common/common_types.h"
+
+namespace Dynarmic::Backend::LoongArch64 {
+
+constexpr size_t SpillCount = 64;
+
+struct alignas(16) StackLayout {
+    s64 cycles_remaining;
+    s64 cycles_to_run;
+
+    std::array<u64, SpillCount> spill;
+
+    u32 save_host_fpcr;
+    u32 save_host_fpsr;
+
+    bool check_bit;
+};
+
+static_assert(sizeof(StackLayout) % 16 == 0);
+
+}  // namespace Dynarmic::Backend::LoongArch64

src/dynarmic/src/dynarmic/common/atomic.h

@@ -36,6 +36,14 @@ inline void And(volatile u32* ptr, u32 value) {
 #endif
 }
 
+inline u32 Exchange(volatile u32* ptr, u32 value) {
+#ifdef _MSC_VER
+    return static_cast<u32>(_InterlockedExchange(reinterpret_cast<volatile long*>(ptr), value));
+#else
+    return __atomic_exchange_n(ptr, value, __ATOMIC_SEQ_CST);
+#endif
+}
+
 inline void Barrier() {
 #ifdef _MSC_VER
     _ReadWriteBarrier();