Avoid copying memory representation of undef data

During MIR interpretation it may happen that a place containing uninitialized bytes is copied. This would read the current representation of these bytes and write it to the destination even though they must, by definition, not matter to the execution. This elides that representation change when no bytes are defined in such a copy, saving some cpu cycles. In such a case, the memory of the target allocation is not touched at all which also means that sometimes no physical page backing the memory allocation of the representation needs to be provided by the OS at all, reducing memory pressure on the system.
rust-lang · 197g · Oct 2, 2019 · Dec 2, 2019 · Dec 2, 2019 · e35d56b971f4e1dda8cde421d6de73d27f7a5b0f
commit e35d56b971f4e1dda8cde421d6de73d27f7a5b0f
diff --git a/src/librustc/mir/interpret/allocation.rs b/src/librustc/mir/interpret/allocation.rs
@@ -695,6 +695,12 @@ impl<Tag, Extra> Allocation<Tag, Extra> {
     }
 }
 
+impl AllocationDefinedness {
+    pub fn all_bytes_undef(&self) -> bool {
+        self.initial == false && self.ranges.len() == 1
-        self.initial == false && self.ranges.len() == 1
+		// The `ranges` are run-length encoded and of alternating definedness.
+		// So if `ranges.len() > 1` then the second block is a range of defined bytes.
+        self.initial == false && self.ranges.len() == 1
-        self.initial == false && self.ranges.len() == 1
+		// The `ranges` are run-length encoded and of alternating definedness.
+		// So if `ranges.len() > 1` then the second block is a range of defined bytes.
+        self.initial == false && self.ranges.len() == 1
+    }
+}
+
 /// Relocations.
 #[derive(Clone, PartialEq, Eq, PartialOrd, Ord, Hash, Debug, RustcEncodable, RustcDecodable)]
 pub struct Relocations<Tag = (), Id = AllocId>(SortedMap<Size, (Tag, Id)>);

diff --git a/src/librustc_mir/interpret/memory.rs b/src/librustc_mir/interpret/memory.rs
@@ -855,6 +855,22 @@ impl<'mir, 'tcx, M: Machine<'mir, 'tcx>> Memory<'mir, 'tcx, M> {
         let relocations = self.get_raw(src.alloc_id)?
             .prepare_relocation_copy(self, src, size, dest, length);
 
+        // Prepare a copy of the undef mask.
+        let compressed = self.get_raw(src.alloc_id)?.compress_undef_range(src, size);
+
+        if compressed.all_bytes_undef() {
+            // Fast path: If all bytes are `undef` then there is nothing to copy. The target range
+            // is marked as undef but we otherwise omit changing the byte representation which may
+            // be arbitrary for undef bytes.
+            // This also avoids writing to the target bytes so that the backing allocation is never
+            // touched if the bytes stay undef for the whole interpreter execution. On contemporary
+            // operating system this can avoid physically allocating the page.
+            let dest_alloc = self.get_raw_mut(dest.alloc_id)?;
+            dest_alloc.mark_definedness(dest, size * length, false);
+            dest_alloc.mark_relocation_range(relocations);
+            return Ok(());
+        }
+
         let tcx = self.tcx.tcx;
 
         // This checks relocation edges on the src.
@@ -897,8 +913,10 @@ impl<'mir, 'tcx, M: Machine<'mir, 'tcx>> Memory<'mir, 'tcx, M> {
             }
         }
 
-        // copy definedness to the destination
-        self.copy_undef_mask(src, dest, size, length)?;
+        // now copy over the undef data
+        self.get_raw_mut(dest.alloc_id)?
+            .mark_compressed_undef_range(&compressed, dest, size, length);
+
         // copy the relocations to the destination
         self.get_raw_mut(dest.alloc_id)?.mark_relocation_range(relocations);
 
@@ -908,27 +926,6 @@ impl<'mir, 'tcx, M: Machine<'mir, 'tcx>> Memory<'mir, 'tcx, M> {
 
 /// Undefined bytes
 impl<'mir, 'tcx, M: Machine<'mir, 'tcx>> Memory<'mir, 'tcx, M> {
-    // FIXME: Add a fast version for the common, nonoverlapping case
-    fn copy_undef_mask(
-        &mut self,
-        src: Pointer<M::PointerTag>,
-        dest: Pointer<M::PointerTag>,
-        size: Size,
-        repeat: u64,
-    ) -> InterpResult<'tcx> {
-        // The bits have to be saved locally before writing to dest in case src and dest overlap.
-        assert_eq!(size.bytes() as usize as u64, size.bytes());
-
-        let src_alloc = self.get_raw(src.alloc_id)?;
-        let compressed = src_alloc.compress_undef_range(src, size);
-
-        // now fill in all the data
-        let dest_allocation = self.get_raw_mut(dest.alloc_id)?;
-        dest_allocation.mark_compressed_undef_range(&compressed, dest, size, repeat);
-
-        Ok(())
-    }
-
     pub fn force_ptr(
         &self,
         scalar: Scalar<M::PointerTag>,