first intent to pinpoint performance decrease on virtual buffer table + multi-range

src/video_core/buffer_cache/virtual_range_cache.h

@@ -5,9 +5,7 @@
 
 #include <atomic>
 #include <limits>
-#include <mutex>
 #include <optional>
-#include <vector>
 
 #include <boost/container/small_vector.hpp>
 
@@ -28,23 +26,22 @@ using VirtualSegments = boost::container::small_vector<VirtualSegment, 8>;
 class VirtualRangeCache {
 public:
     static constexpr size_t MAX_ENTRIES = 8192;
-    static constexpr size_t MAX_DEFERRED = 4096;
 
     const VirtualSegments* Query(Tegra::MemoryManager& memory, GPUVAddr gpu_addr, u32 size) {
-        if (has_deferred.load(std::memory_order_acquire)) {
-            ApplyDeferred();
-        }
+        const u64 current = generation.load(std::memory_order_acquire);
         if (entries.size() > MAX_ENTRIES) {
             entries.clear();
         }
         const size_t as_id = memory.GetID();
         const u64 key = MakeKey(as_id, gpu_addr);
         const auto it = entries.find(key);
-        if (it != entries.end() && it->second.as_id == as_id &&
-            it->second.gpu_addr == gpu_addr && it->second.size == size) {
+        if (it != entries.end() && it->second.generation == current &&
+            it->second.as_id == as_id && it->second.gpu_addr == gpu_addr &&
+            it->second.size == size) {
             return &it->second.segments;
         }
         Entry entry{};
+        entry.generation = current;
         entry.as_id = as_id;
         entry.gpu_addr = gpu_addr;
         entry.size = size;
@@ -79,95 +76,27 @@ public:
         return &result.first->second.segments;
     }
 
-    void Unmap(size_t as_id, GPUVAddr gpu_addr, u64 size) {
-        if (size == 0) {
-            return;
+    void Unmap(size_t, GPUVAddr, u64 size) {
+        if (size != 0) {
+            generation.fetch_add(1, std::memory_order_release);
         }
-        {
-            std::scoped_lock lock{deferred_mutex};
-            if (!deferred.empty()) {
-                DeferredUnmap& last = deferred.back();
-                if (last.as_id == as_id && last.gpu_addr + last.size == gpu_addr) {
-                    last.size += size;
-                    has_deferred.store(true, std::memory_order_release);
-                    return;
-                }
-            }
-            if (deferred.size() >= MAX_DEFERRED) {
-                deferred.clear();
-                deferred_overflow = true;
-            } else {
-                deferred.push_back(DeferredUnmap{
-                    .as_id = as_id,
-                    .gpu_addr = gpu_addr,
-                    .size = size,
-                });
-            }
-        }
-        has_deferred.store(true, std::memory_order_release);
     }
 
 private:
     struct Entry {
         VirtualSegments segments;
+        u64 generation{};
         size_t as_id{};
         GPUVAddr gpu_addr{};
         u32 size{};
     };
 
-    struct DeferredUnmap {
-        size_t as_id;
-        GPUVAddr gpu_addr;
-        u64 size;
-    };
-
     static u64 MakeKey(size_t as_id, GPUVAddr gpu_addr) {
         return (static_cast<u64>(as_id) << 48) ^ gpu_addr;
     }
 
-    void ApplyDeferred() {
-        std::vector<DeferredUnmap> pending;
-        bool overflow = false;
-        {
-            std::scoped_lock lock{deferred_mutex};
-            has_deferred.store(false, std::memory_order_release);
-            pending.swap(deferred);
-            overflow = deferred_overflow;
-            deferred_overflow = false;
-        }
-        if (overflow) {
-            entries.clear();
-            return;
-        }
-        if (pending.empty() || entries.empty()) {
-            return;
-        }
-        for (auto it = entries.begin(); it != entries.end();) {
-            const Entry& entry = it->second;
-            const GPUVAddr entry_end = entry.gpu_addr + entry.size;
-            bool overlaps = false;
-            for (const DeferredUnmap& unmap : pending) {
-                if (unmap.as_id != entry.as_id) {
-                    continue;
-                }
-                if (entry.gpu_addr < unmap.gpu_addr + unmap.size && unmap.gpu_addr < entry_end) {
-                    overlaps = true;
-                    break;
-                }
-            }
-            if (overlaps) {
-                it = entries.erase(it);
-            } else {
-                ++it;
-            }
-        }
-    }
-
     ::Common::unordered_map<u64, Entry> entries;
-    std::vector<DeferredUnmap> deferred;
-    std::mutex deferred_mutex;
-    std::atomic<bool> has_deferred{false};
-    bool deferred_overflow{};
+    std::atomic<u64> generation{};
 };
 
 } // namespace VideoCommon