[TEST] Sync and copy buffer hazards

src/video_core/buffer_cache/buffer_cache.h

@@ -207,12 +207,22 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     std::optional<WindowBufferRef> src_window;
     std::optional<WindowBufferRef> dst_window;
     if constexpr (USE_UNIFIED_DIRECT_BINDING) {
-        src_window = TryObtainWindowBuffer(*cpu_src_address, static_cast<u32>(amount),
-                                           UNIFIED_COPY_BINDING_ALIGNMENT,
-                                           ObtainBufferOperation::DoNothing);
-        dst_window = TryObtainWindowBuffer(*cpu_dest_address, static_cast<u32>(amount),
-                                           UNIFIED_COPY_BINDING_ALIGNMENT,
-                                           ObtainBufferOperation::DoNothing);
+        if (amount <= (std::numeric_limits<u32>::max)()) {
+            src_window = TryObtainWindowBuffer(*cpu_src_address, static_cast<u32>(amount),
+                                               UNIFIED_COPY_BINDING_ALIGNMENT,
+                                               ObtainBufferOperation::DoNothing);
+            dst_window = TryObtainWindowBuffer(*cpu_dest_address, static_cast<u32>(amount),
+                                               UNIFIED_COPY_BINDING_ALIGNMENT,
+                                               ObtainBufferOperation::DoNothing);
+            if (src_window && dst_window &&
+                src_window->window_index == dst_window->window_index) {
+                const u64 src_begin = src_window->offset;
+                const u64 dst_begin = dst_window->offset;
+                if (src_begin < dst_begin + amount && dst_begin < src_begin + amount) {
+                    dst_window.reset();
+                }
+            }
+        }
     }
 
     ClearDownload(*cpu_dest_address, amount);
@@ -364,11 +374,10 @@ void BufferCache<P>::MarkGpuWrittenRange(DAddr device_addr, u32 size) {
     if (size == 0) {
         return;
     }
-    const BufferId buffer_id = FindBuffer(device_addr, size);
-    if (buffer_id == NULL_BUFFER_ID) {
-        return;
-    }
-    MarkWrittenBuffer(buffer_id, device_addr, size);
+    memory_tracker.MarkRegionAsGpuModified(device_addr, size);
+    gpu_modified_ranges.Add(device_addr, size);
+    uncommitted_gpu_modified_ranges.Add(device_addr, size);
+    in_place_gpu_written_ranges.Subtract(device_addr, size);
 }
 
 template <class P>
@@ -1868,14 +1877,6 @@ void BufferCache<P>::ForgetGpuModifiedRange(DAddr device_addr, u64 size) {
     in_place_gpu_written_ranges.Subtract(device_addr, size);
 }
 
-template <class P>
-bool BufferCache<P>::IsRegionWrittenInPlace(DAddr device_addr, u64 size) {
-    u64 covered = 0;
-    in_place_gpu_written_ranges.ForEachInRange(
-        device_addr, size, [&](DAddr start, DAddr end) { covered += end - start; });
-    return covered == size;
-}
-
 template <class P>
 bool BufferCache<P>::IntersectsInPlaceWrites(DAddr device_addr, u64 size) {
     bool intersects = false;
@@ -2212,8 +2213,7 @@ bool BufferCache<P>::ResolveUnifiedDirectBinding([[maybe_unused]] DAddr device_a
             }
             contiguous += Core::DEVICE_PAGESIZE;
         }
-        if (IsRegionGpuModified(device_addr, size) &&
-            !IsRegionWrittenInPlace(device_addr, size)) {
+        if (IsRegionGpuModified(device_addr, size)) {
             return false;
         }
         window_index = relative / window_size;

src/video_core/buffer_cache/buffer_cache_base.h

@@ -10,6 +10,7 @@
 #include <array>
 #include <bit>
 #include <functional>
+#include <limits>
 #include <memory>
 #include <mutex>
 #include <numeric>
@@ -449,8 +450,6 @@ private:
 
     void ForgetGpuModifiedRange(DAddr device_addr, u64 size);
 
-    [[nodiscard]] bool IsRegionWrittenInPlace(DAddr device_addr, u64 size);
-
     [[nodiscard]] bool IntersectsInPlaceWrites(DAddr device_addr, u64 size);
 
     void WaitForInPlaceWrites();