feat: improve performance by reducing the trashing of zlib::Inflate.

Previously, for every zlib inflate operation, we would allocate memory
due to the creation of a decompressor.

This is now avoided by reusing thread-local copies of such a decompressor,
which is reused and reset instead.

Author: Byron

gix-pack/src/cache/delta/traverse/resolve.rs

@@ -405,11 +405,19 @@ fn set_len(v: &mut Vec<u8>, new_len: usize) {
 
 fn decompress_all_at_once_with(b: &[u8], decompressed_len: usize, out: &mut Vec<u8>) -> Result<(), Error> {
     set_len(out, decompressed_len);
-    zlib::Inflate::default()
-        .once(b, out)
-        .map_err(|err| Error::ZlibInflate {
+    use std::cell::RefCell;
+    thread_local! {
+        pub static INFLATE: RefCell<zlib::Inflate> = RefCell::new(zlib::Inflate::default());
+    }
+
+    INFLATE.with(|inflate| {
+        let mut inflate = inflate.borrow_mut();
+        inflate.reset();
+        let res = inflate.once(b, out).map_err(|err| Error::ZlibInflate {
             source: err,
             message: "Failed to decompress entry",
-        })?;
+        });
+        res
+    })?;
     Ok(())
 }

gix-pack/src/data/file/decode/entry.rs

@@ -126,9 +126,18 @@ impl File {
         let offset: usize = data_offset.try_into().expect("offset representable by machine");
         assert!(offset < self.data.len(), "entry offset out of bounds");
 
-        zlib::Inflate::default()
-            .once(&self.data[offset..], out)
-            .map(|(_status, consumed_in, _consumed_out)| consumed_in)
+        use std::cell::RefCell;
+        thread_local! {
+            pub static INFLATE: RefCell<zlib::Inflate> = RefCell::new(zlib::Inflate::default());
+        }
+        INFLATE.with(|inflate| {
+            let mut inflate = inflate.borrow_mut();
+            let res = inflate
+                .once(&self.data[offset..], out)
+                .map(|(_status, consumed_in, _consumed_out)| consumed_in);
+            inflate.reset();
+            res
+        })
     }
 
     /// Like `decompress_entry_from_data_offset`, but returns consumed input and output.
@@ -140,9 +149,19 @@ impl File {
         let offset: usize = data_offset.try_into().expect("offset representable by machine");
         assert!(offset < self.data.len(), "entry offset out of bounds");
 
-        zlib::Inflate::default()
-            .once(&self.data[offset..], out)
-            .map(|(_status, consumed_in, consumed_out)| (consumed_in, consumed_out))
+        use std::cell::RefCell;
+        thread_local! {
+            pub static INFLATE: RefCell<zlib::Inflate> = RefCell::new(zlib::Inflate::default());
+        }
+
+        INFLATE.with(|inflate| {
+            let mut inflate = inflate.borrow_mut();
+            let res = inflate
+                .once(&self.data[offset..], out)
+                .map(|(_status, consumed_in, consumed_out)| (consumed_in, consumed_out));
+            inflate.reset();
+            res
+        })
     }
 
     /// Decode an entry, resolving delta's as needed, while growing the `out` vector if there is not enough