refactor: refactor VU1

ps2xIOP/src/builtin_profiles.cpp

@@ -70,7 +70,8 @@ namespace ps2x::iop::detail
                 .responseCounterOffset = 4u,
                 .zeroReceiveBuffer = true,
                 .signalNowaitCompletion = true,
-                .suppressedCompletionCallbacks = {0x001FFD70u},
+                .completeQueuedPlayStreams = true,
+                .suppressedCompletionCallbacks = {},
             };
         }
 

ps2xIOP/src/module_factories.h

@@ -107,6 +107,7 @@ namespace ps2x::iop::detail
         uint32_t responseCounterOffset = 0u;
         bool zeroReceiveBuffer = true;
         bool signalNowaitCompletion = false;
+        bool completeQueuedPlayStreams = false;
         std::vector<uint32_t> suppressedCompletionCallbacks;
     };
 

ps2xIOP/src/modules/sound_update_stub.cpp

@@ -7,11 +7,20 @@
 #include <stdexcept>
 #include <unordered_set>
 #include <utility>
+#include <vector>
 
 namespace ps2x::iop::detail
 {
     namespace
     {
+        constexpr uint16_t kPlayStreamCommand = 1u;
+        constexpr uint32_t kResponseRecordStride = 0x20u;
+        constexpr uint32_t kPackedStreamOffset = 4u;
+        constexpr uint32_t kStreamSlotMask = 0x3Fu;
+        constexpr uint32_t kStreamSlotCount = 48u;
+        constexpr uint32_t kCommandStreamSlotShift = 8u;
+        constexpr uint32_t kResponseStreamSlotShift = 4u;
+
         class SoundUpdateStubService final : public IopService
         {
         public:
@@ -34,6 +43,7 @@ namespace ps2x::iop::detail
             {
                 std::lock_guard<std::mutex> lock(m_mutex);
                 m_updateCounter = 0u;
+                m_completedStreamCount = 0u;
             }
 
             [[nodiscard]] RpcResult handleRpc(const RpcRequest &request) override
@@ -62,24 +72,44 @@ namespace ps2x::iop::detail
                     (void)m_host.zeroGuest(request.receive.address, request.receive.size);
                 }
 
+                std::vector<uint32_t> activeStreamSlots;
+                if (m_bindings.completeQueuedPlayStreams && request.receive.address != 0u)
+                {
+                    // PlayStream leaves the EE slot in state 2. One active record moves it
+                    // to state 1; the following empty update lets SOUND_CopyIOPBuffer clear it.
+                    activeStreamSlots = findQueuedPlayStreams(request);
+                    trimToReceiveCapacity(activeStreamSlots, request.receive.size);
+                }
+
                 uint32_t counter = 0u;
                 {
                     std::lock_guard<std::mutex> lock(m_mutex);
                     counter = ++m_updateCounter;
+                    m_completedStreamCount += activeStreamSlots.size();
                 }
 
-                constexpr uint32_t activeStreams = 0u;
+                const uint32_t activeStreams = static_cast<uint32_t>(activeStreamSlots.size());
                 if (request.receive.address != 0u &&
                     request.receive.size >= m_bindings.activeStreamCountOffset + sizeof(activeStreams))
                 {
                     const uint32_t address = request.receive.address + m_bindings.activeStreamCountOffset;
                     (void)m_host.writeGuest(address, &activeStreams, sizeof(activeStreams));
                 }
 
+                for (size_t index = 0u; index < activeStreamSlots.size(); ++index)
+                {
+                    const uint32_t packedStream = activeStreamSlots[index] << kResponseStreamSlotShift;
+                    const uint32_t offset = m_bindings.activeStreamCountOffset + static_cast<uint32_t>(index) * kResponseRecordStride + kPackedStreamOffset;
+                    const uint32_t address = request.receive.address + offset;
+                    (void)m_host.writeGuest(address, &packedStream, sizeof(packedStream));
+                }
+
+                const uint32_t counterOffset = m_bindings.responseCounterOffset +
+                                               activeStreams * kResponseRecordStride;
                 if (request.receive.address != 0u &&
-                    request.receive.size >= m_bindings.responseCounterOffset + sizeof(counter))
+                    request.receive.size >= counterOffset + sizeof(counter))
                 {
-                    const uint32_t address = request.receive.address + m_bindings.responseCounterOffset;
+                    const uint32_t address = request.receive.address + counterOffset;
                     (void)m_host.writeGuest(address, &counter, sizeof(counter));
                 }
 
@@ -90,14 +120,98 @@ namespace ps2x::iop::detail
             {
                 std::lock_guard<std::mutex> lock(m_mutex);
                 metrics.push_back({"update_counter", m_updateCounter, false});
+                metrics.push_back({"completed_streams", m_completedStreamCount, false});
             }
 
         private:
+            [[nodiscard]] std::vector<uint32_t> findQueuedPlayStreams(const RpcRequest &request) const
+            {
+                std::vector<uint32_t> slots;
+                if (request.send.address == 0u || request.send.size < sizeof(uint16_t))
+                {
+                    return slots;
+                }
+
+                uint16_t commandCount = 0u;
+                if (!m_host.readGuest(request.send.address, &commandCount, sizeof(commandCount)))
+                {
+                    return slots;
+                }
+
+                uint32_t offset = sizeof(commandCount);
+                for (uint32_t commandIndex = 0u; commandIndex < commandCount; ++commandIndex)
+                {
+                    constexpr uint32_t headerSize = sizeof(uint16_t) * 2u;
+                    if (offset > request.send.size || request.send.size - offset < headerSize)
+                    {
+                        break;
+                    }
+
+                    std::array<uint16_t, 2> header{};
+                    if (!m_host.readGuest(request.send.address + offset,
+                                          header.data(),
+                                          sizeof(header)))
+                    {
+                        break;
+                    }
+                    offset += headerSize;
+
+                    const uint32_t argumentBytes =
+                        static_cast<uint32_t>(header[1]) * sizeof(uint16_t);
+                    if (argumentBytes > request.send.size - offset)
+                    {
+                        break;
+                    }
+
+                    if (header[0] == kPlayStreamCommand && header[1] >= 2u)
+                    {
+                        uint16_t encodedSlot = 0u;
+                        if (m_host.readGuest(request.send.address + offset + sizeof(uint16_t),
+                                             &encodedSlot,
+                                             sizeof(encodedSlot)))
+                        {
+                            const uint32_t slot =
+                                (encodedSlot >> kCommandStreamSlotShift) & kStreamSlotMask;
+                            if (slot < kStreamSlotCount &&
+                                std::find(slots.begin(), slots.end(), slot) == slots.end())
+                            {
+                                slots.push_back(slot);
+                            }
+                        }
+                    }
+
+                    offset += argumentBytes;
+                }
+                return slots;
+            }
+
+            void trimToReceiveCapacity(std::vector<uint32_t> &slots, uint32_t receiveSize) const
+            {
+                size_t count = 0u;
+                for (; count < slots.size(); ++count)
+                {
+                    const uint64_t recordOffset =
+                        static_cast<uint64_t>(m_bindings.activeStreamCountOffset) +
+                        static_cast<uint64_t>(count) * kResponseRecordStride +
+                        kPackedStreamOffset;
+                    const uint64_t counterOffset =
+                        static_cast<uint64_t>(m_bindings.responseCounterOffset) +
+                        static_cast<uint64_t>(count + 1u) * kResponseRecordStride;
+                    if (recordOffset + sizeof(uint32_t) > receiveSize ||
+                        counterOffset + sizeof(uint32_t) > receiveSize)
+                    {
+                        break;
+                    }
+                }
+                slots.resize(count);
+            }
+
             IopHost &m_host;
             SoundUpdateStubBindings m_bindings;
             std::array<uint32_t, 1> m_sids;
             mutable std::mutex m_mutex;
             uint32_t m_updateCounter = 0u;
+            uint64_t m_completedStreamCount = 0u;
         };
     }
 

ps2xRecomp/include/ps2recomp/instructions.h

@@ -670,37 +670,21 @@ namespace ps2recomp
     //     VU0_VLDQ = 0x1F    // VU0 Load/Store Quad with Decrement
     // };
 
-    // VU0 Control Register Numbers (used with CFC2/CTC2)
+    // VU0 COP2 control register numbers used by CFC2/CTC2.
+    // Registers 0..15 address VI0..VI15 directly.
     enum VU0ControlRegisters
     {
-        VU0_CR_STATUS = 0, // Status/Control register
-        VU0_CR_MAC = 1,    // MAC flags register
-        VU0_CR_CLIP = 5,   // Clipping flags register
-        VU0_CR_R = 3,      // R register (Random number)
-        VU0_CR_I = 4,      // I register (Immediate)
-
-        // Add missing registers
-        VU0_CR_VPU_STAT = 2,   // VPU-STAT register
-        VU0_CR_TPC = 6,        // T (program counter) register
-        VU0_CR_CMSAR0 = 7,     // Call/return address 0
-        VU0_CR_FBRST = 8,      // VIF/VU reset register
-        VU0_CR_VPU_STAT2 = 9,  // VPU-STAT register 2
-        VU0_CR_TPC2 = 10,      // T (program counter) register 2
-        VU0_CR_CMSAR1 = 11,    // Call/return address 1
-        VU0_CR_FBRST2 = 12,    // VIF/VU reset register 2
-        VU0_CR_VPU_STAT3 = 13, // VPU-STAT register 3
-        VU0_CR_CMSAR2 = 14,    // Call/return address 2
-        VU0_CR_FBRST3 = 15,    // VIF/VU reset register 3
-        VU0_CR_VPU_STAT4 = 16, // VPU-STAT register 4
-        VU0_CR_CMSAR3 = 17,    // Call/return address 3
-        VU0_CR_FBRST4 = 18,    // VIF/VU reset register 4
-        VU0_CR_ACC = 20,       // Accumulator register
-        VU0_CR_INFO = 21,      // Information register
-        VU0_CR_CLIP2 = 22,     // Clipping flags register 2
-        VU0_CR_P = 26,         // P register
-        VU0_CR_XITOP = 27,     // XITOP register
-        VU0_CR_ITOP = 28,      // ITOP register
-        VU0_CR_TOP = 29        // TOP register
+        VU0_CR_STATUS = 16,
+        VU0_CR_MAC = 17,
+        VU0_CR_CLIP = 18,
+        VU0_CR_R = 20,
+        VU0_CR_I = 21,
+        VU0_CR_Q = 22,
+        VU0_CR_TPC = 26,
+        VU0_CR_CMSAR0 = 27,
+        VU0_CR_FBRST = 28,
+        VU0_CR_VPU_STAT = 29,
+        VU0_CR_CMSAR1 = 31
     };
     enum VU0OPSFunctions
     {

ps2xRecomp/include/ps2recomp/ps2_recompiler.h

@@ -34,8 +34,10 @@ namespace ps2recomp
         bool recompile();
         void generateOutput();
         void printReport() const;
+        const RecompilerReporter::Counters &reportCounters() const { return m_reporter.counters(); }
 
         static StubTarget resolveStubTarget(const std::string& name);
+        static bool IsCorrectnessCriticalFunctionName(const std::string &name);
         static size_t DiscoverAdditionalEntryPoints(
             std::vector<Function> &functions,
             std::unordered_map<uint32_t, std::vector<Instruction>> &decodedFunctions,
@@ -65,6 +67,7 @@ namespace ps2recomp
         std::unordered_set<std::string> m_stubFunctions;
         std::unordered_set<uint32_t> m_stubFunctionStarts;
         std::unordered_map<uint32_t, std::string> m_stubHandlerBindingsByStart;
+        std::unordered_set<uint32_t> m_correctnessCriticalFunctionStarts;
         std::map<uint32_t, std::string> m_generatedStubs;
         std::unordered_map<uint32_t, std::string> m_functionRenames;
         std::unordered_map<uint32_t, std::vector<uint32_t>> m_resumeEntryTargetsByOwner;
@@ -74,6 +77,9 @@ namespace ps2recomp
         void discoverAdditionalEntryPoints();
         bool shouldSkipFunction(const Function &function) const;
         bool isStubFunction(const Function &function) const;
+        bool isCorrectnessCriticalFunction(const Function &function) const;
+        bool hasResolvedStubHandler(const Function &function) const;
+        void collectCorrectnessCriticalFunctionStarts();
         bool generateFunctionHeader();
         bool generateStubHeader();
         bool writeToFile(const std::string &path, const std::string &content);

ps2xRecomp/include/ps2recomp/recompiler_reporter.h

@@ -46,6 +46,8 @@ namespace ps2recomp
             size_t unhandledInstructions = 0;
             size_t indirectFallbackPromotions = 0;
             size_t indirectFallbackEntries = 0;
+            size_t correctnessCriticalGuestFallbacks = 0;
+            size_t correctnessCriticalFailures = 0;
         };
 
         void progress(const std::string &message);
@@ -63,6 +65,8 @@ namespace ps2recomp
         void recordDecodeFailure();
         void recordAdditionalEntryPoints(size_t count);
         void recordGeneratedFunctions(size_t count);
+        void recordCorrectnessCriticalGuestFallback();
+        void recordCorrectnessCriticalFailure();
         void recordIndirectFallbackPromotion(const std::string &functionName,
                                              const std::vector<uint32_t> &jumpAddresses,
                                              size_t promotedEntryCount);