perf: use subarray instead of slice in dispatch minimal (#4180)

4 files changed, 36 insertions, 26 deletions

diff --git a/cli/js/dispatch_minimal.ts b/cli/js/dispatch_minimal.ts
index 5778b1333..567dcc963 100644
--- a/cli/js/dispatch_minimal.ts
+++ b/cli/js/dispatch_minimal.ts
@@ -27,7 +27,7 @@ export interface RecordMinimal {
 }
 
 export function recordFromBufMinimal(ui8: Uint8Array): RecordMinimal {
-  const header = ui8.slice(0, 12);
+  const header = ui8.subarray(0, 12);
   const buf32 = new Int32Array(
     header.buffer,
     header.byteOffset,
@@ -40,7 +40,7 @@ export function recordFromBufMinimal(ui8: Uint8Array): RecordMinimal {
 
   if (arg < 0) {
     const kind = result as ErrorKind;
-    const message = decoder.decode(ui8.slice(12));
+    const message = decoder.decode(ui8.subarray(12));
     err = { kind, message };
   } else if (ui8.length != 12) {
     throw new errors.InvalidData("BadMessage");
diff --git a/core/shared_queue.js b/core/shared_queue.js
index 093cc223f..742a90995 100644
--- a/core/shared_queue.js
+++ b/core/shared_queue.js
@@ -115,7 +115,8 @@ SharedQueue Binary Layout
       if (index == 0) {
         return HEAD_INIT;
       } else {
-        return shared32[INDEX_OFFSETS + 2 * (index - 1)];
+        const prevEnd = shared32[INDEX_OFFSETS + 2 * (index - 1)];
+        return (prevEnd + 3) & ~3;
       }
     } else {
       return null;
@@ -125,16 +126,18 @@ SharedQueue Binary Layout
   function push(opId, buf) {
     const off = head();
     const end = off + buf.byteLength;
+    const alignedEnd = (end + 3) & ~3;
     const index = numRecords();
-    if (end > shared32.byteLength || index >= MAX_RECORDS) {
+    if (alignedEnd > shared32.byteLength || index >= MAX_RECORDS) {
       // console.log("shared_queue.js push fail");
       return false;
     }
     setMeta(index, end, opId);
+    assert(alignedEnd % 4 === 0);
     assert(end - off == buf.byteLength);
     sharedBytes.set(buf, off);
     shared32[INDEX_NUM_RECORDS] += 1;
-    shared32[INDEX_HEAD] = end;
+    shared32[INDEX_HEAD] = alignedEnd;
     return true;
   }
 
diff --git a/core/shared_queue.rs b/core/shared_queue.rs
index dad3e380d..0f35d1310 100644
--- a/core/shared_queue.rs
+++ b/core/shared_queue.rs
@@ -134,7 +134,8 @@ impl SharedQueue {
         HEAD_INIT
       } else {
         let s = self.as_u32_slice();
-        s[INDEX_OFFSETS + 2 * (index - 1)] as usize
+        let prev_end = s[INDEX_OFFSETS + 2 * (index - 1)] as usize;
+        (prev_end + 3) & !3
       })
     } else {
       None
@@ -153,7 +154,6 @@ impl SharedQueue {
 
     let off = self.get_offset(i).unwrap();
     let (op_id, end) = self.get_meta(i).unwrap();
-
     if self.size() > 1 {
       let u32_slice = self.as_u32_slice_mut();
       u32_slice[INDEX_NUM_SHIFTED_OFF] += 1;
@@ -169,29 +169,33 @@ impl SharedQueue {
     Some((op_id, &self.bytes()[off..end]))
   }
 
-  /// Because JS-side may cast `record` to Int32Array it is required
-  /// that `record`'s length is divisible by 4.
+  /// Because JS-side may cast popped message to Int32Array it is required
+  /// that every message is aligned to 4-bytes.
   pub fn push(&mut self, op_id: OpId, record: &[u8]) -> bool {
     let off = self.head();
+    assert_eq!(off % 4, 0);
     let end = off + record.len();
+    let aligned_end = (end + 3) & !3;
     debug!(
-      "rust:shared_queue:pre-push: op={}, off={}, end={}, len={}",
+      "rust:shared_queue:pre-push: op={}, off={}, end={}, len={}, aligned_end={}",
       op_id,
       off,
       end,
-      record.len()
+      record.len(),
+      aligned_end,
     );
     let index = self.num_records();
-    if end > self.bytes().len() || index >= MAX_RECORDS {
+    if aligned_end > self.bytes().len() || index >= MAX_RECORDS {
       debug!("WARNING the sharedQueue overflowed");
       return false;
     }
+    assert_eq!(aligned_end % 4, 0);
     self.set_meta(index, end, op_id);
     assert_eq!(end - off, record.len());
     self.bytes_mut()[off..end].copy_from_slice(record);
     let u32_slice = self.as_u32_slice_mut();
     u32_slice[INDEX_NUM_RECORDS] += 1;
-    u32_slice[INDEX_HEAD] = end as u32;
+    u32_slice[INDEX_HEAD] = aligned_end as u32;
     debug!(
       "rust:shared_queue:push: num_records={}, num_shifted_off={}, head={}",
       self.num_records(),
@@ -258,15 +262,15 @@ mod tests {
   #[test]
   fn overflow() {
     let mut q = SharedQueue::new(RECOMMENDED_SIZE);
-    assert!(q.push(0, &alloc_buf(RECOMMENDED_SIZE - 1)));
+    assert!(q.push(0, &alloc_buf(RECOMMENDED_SIZE - 5)));
     assert_eq!(q.size(), 1);
-    assert!(!q.push(0, &alloc_buf(2)));
+    assert!(!q.push(0, &alloc_buf(6)));
     assert_eq!(q.size(), 1);
     assert!(q.push(0, &alloc_buf(1)));
     assert_eq!(q.size(), 2);
 
     let (_op_id, buf) = q.shift().unwrap();
-    assert_eq!(buf.len(), RECOMMENDED_SIZE - 1);
+    assert_eq!(buf.len(), RECOMMENDED_SIZE - 5);
     assert_eq!(q.size(), 1);
 
     assert!(!q.push(0, &alloc_buf(1)));
@@ -291,14 +295,13 @@ mod tests {
   #[test]
   fn allow_any_buf_length() {
     let mut q = SharedQueue::new(RECOMMENDED_SIZE);
-    // check that `record` that has length not a multiple of 4 will cause panic
-    q.push(0, &alloc_buf(1));
-    q.push(0, &alloc_buf(2));
-    q.push(0, &alloc_buf(3));
-    q.push(0, &alloc_buf(4));
-    q.push(0, &alloc_buf(5));
-    q.push(0, &alloc_buf(6));
-    q.push(0, &alloc_buf(7));
-    q.push(0, &alloc_buf(8));
+    // Check that `record` that has length not a multiple of 4 will
+    // not cause panic. Still make sure that records are always
+    // aligned to 4 bytes.
+    for i in 1..9 {
+      q.push(0, &alloc_buf(i));
+      assert_eq!(q.num_records(), i);
+      assert_eq!(q.head() % 4, 0);
+    }
   }
 }
diff --git a/core/shared_queue_test.js b/core/shared_queue_test.js
index 37bc9a28a..7821f93a6 100644
--- a/core/shared_queue_test.js
+++ b/core/shared_queue_test.js
@@ -28,10 +28,14 @@ function main() {
   const h = q.head();
   assert(h > 0);
 
+  // This record's len is not divisble by
+  // 4 so after pushing it to the queue,
+  // next record offset should be aligned to 4.
   let r = new Uint8Array([1, 2, 3, 4, 5]);
   const len = r.byteLength + h;
   assert(q.push(1, r));
-  assert(q.head() == len);
+  // Record should be aligned to 4 bytes
+  assert(q.head() == len + 3);
 
   r = new Uint8Array([6, 7]);
   assert(q.push(1, r));