Revert "Draw: Avoid temporary copy for mesh triangulation index buffer"

This reverts commit 108ab1df2d.

This causes issues when duplicating objects that I don't have time
to investigate right now.

source/blender/draw/intern/mesh_extractors/extract_mesh_ibo_tris.cc

@@ -23,14 +23,7 @@ static void extract_tris_mesh(const MeshRenderData &mr,
   const Span<int3> corner_tris = mr.mesh->corner_tris();
   if (!face_sorted.face_tri_offsets) {
     /* There are no hidden faces and no reordering is necessary to group triangles with the same
-     * material. The corner indices from #Mesh::corner_tris() can be copied directly to the GPU
-     * without the usual CPU-side copy owned by the index buffer. Crucially, this assumes that the
-     * data is uploaded to the GPU *before* the dependency graph's evaluated state is cleared (and
-     * with it, the evaluated mesh's triangulation data).
-     *
-     * Eventually these local copies should be completely removed, and code should write directly
-     * to GPU memory, but even then it could be helpful to know that the data already exists
-     * contiguously, owned elsewhere by Blender. */
+     * material. The corner indices from #Mesh::corner_tris() can be copied directly to the GPU. */
     BLI_assert(face_sorted.visible_tris_num == corner_tris.size());
     GPU_indexbuf_build_in_place_from_memory(&ibo,
                                             GPU_PRIM_TRIS,

source/blender/gpu/GPU_index_buffer.hh

@@ -53,8 +53,6 @@ class IndexBuf {
   bool is_subrange_ = false;
   /** True if buffer only contains restart indices. */
   bool is_empty_ = false;
-  /** #data_ is read-only, not owned by an index buffer. */
-  bool reference_data_ = false;
 
   union {
     /** Mapped buffer data. non-NULL indicates not yet sent to VRAM. */
@@ -72,8 +70,7 @@ class IndexBuf {
             uint min_index,
             uint max_index,
             GPUPrimType prim_type,
-            bool uses_restart_indices,
-            bool reference_data);
+            bool uses_restart_indices);
   void init_subrange(IndexBuf *elem_src, uint start, uint length);
   void init_build_on_device(uint index_len);
 

source/blender/gpu/intern/gpu_index_buffer.cc

@@ -315,15 +315,13 @@ void IndexBuf::init(uint indices_len,
                     uint min_index,
                     uint max_index,
                     GPUPrimType prim_type,
-                    bool uses_restart_indices,
-                    bool reference_data)
+                    bool uses_restart_indices)
 {
   is_init_ = true;
   data_ = indices;
   index_start_ = 0;
   index_len_ = indices_len;
   is_empty_ = min_index > max_index;
-  reference_data_ = reference_data;
 
   /* Patch index buffer to remove restart indices from
    * non-restart-compatible primitive types. Restart indices
@@ -494,8 +492,7 @@ void GPU_indexbuf_build_in_place(GPUIndexBufBuilder *builder, IndexBuf *elem)
              builder->index_min,
              builder->index_max,
              builder->prim_type,
-             builder->uses_restart_indices,
-             false);
+             builder->uses_restart_indices);
   builder->data = nullptr;
 }
 
@@ -513,8 +510,7 @@ void GPU_indexbuf_build_in_place_ex(GPUIndexBufBuilder *builder,
              index_min,
              index_max,
              builder->prim_type,
-             uses_restart_indices,
-             false);
+             uses_restart_indices);
   builder->data = nullptr;
 }
 
@@ -526,16 +522,15 @@ void GPU_indexbuf_build_in_place_from_memory(IndexBuf *ibo,
                                              const int32_t index_max,
                                              const bool uses_restart_indices)
 {
-  /* If restart indices are used, they need to be stripped on Metal which would require a copy. */
-  BLI_assert(!uses_restart_indices);
   const uint32_t indices_num = data_len * indices_per_primitive(prim_type);
-  ibo->init(indices_num,
-            const_cast<uint32_t *>(data),
-            index_min,
-            index_max,
-            prim_type,
-            uses_restart_indices,
-            true);
+  /* TODO: The need for this copy is meant to be temporary. The data should be uploaded directly to
+   * the GPU here rather than copied to an array owned by the IBO first. */
+  uint32_t *copy = static_cast<uint32_t *>(
+      MEM_malloc_arrayN(indices_num, sizeof(uint32_t), __func__));
+  threading::memory_bandwidth_bound_task(sizeof(uint32_t) * indices_num * 2, [&]() {
+    array_utils::copy(Span(data, indices_num), MutableSpan(copy, indices_num));
+  });
+  ibo->init(indices_num, copy, index_min, index_max, prim_type, uses_restart_indices);
 }
 
 void GPU_indexbuf_create_subrange_in_place(IndexBuf *elem,

source/blender/gpu/metal/mtl_index_buffer.mm

@@ -139,10 +139,7 @@ void MTLIndexBuf::upload_data()
     }
 
     /* No need to keep copy of data_ in system memory. */
-    if (reference_data_) {
-      data_ = nullptr;
-    }
-    else {
+    if (data_) {
       MEM_SAFE_FREE(data_);
     }
   }

source/blender/gpu/opengl/gl_index_buffer.cc

@@ -36,12 +36,7 @@ void GLIndexBuf::bind()
     /* Sends data to GPU. */
     glBufferData(GL_ELEMENT_ARRAY_BUFFER, size, data_, GL_STATIC_DRAW);
     /* No need to keep copy of data in system memory. */
-    if (reference_data_) {
-      data_ = nullptr;
-    }
-    else {
-      MEM_SAFE_FREE(data_);
-    }
+    MEM_SAFE_FREE(data_);
   }
 }
 

source/blender/gpu/vulkan/vk_index_buffer.cc

@@ -33,12 +33,7 @@ void VKIndexBuffer::ensure_updated()
   VKStagingBuffer staging_buffer(buffer_, VKStagingBuffer::Direction::HostToDevice);
   staging_buffer.host_buffer_get().update(data_);
   staging_buffer.copy_to_device(context);
-  if (reference_data_) {
-    data_ = nullptr;
-  }
-  else {
-    MEM_SAFE_FREE(data_);
-  }
+  MEM_SAFE_FREE(data_);
 }
 
 void VKIndexBuffer::upload_data()