diff --git a/src/index.ts b/src/index.ts
index f1d0d07..1cad6bf 100644
--- a/src/index.ts
+++ b/src/index.ts
@@ -662,7 +662,10 @@ export function decompress(dat: Uint8Array, buf?: Uint8Array, dic?: Uint8Array)
       if (dic) rdic(dic, st);
       if (nb) {
         buf = null;
-        if (st.w.length == st.u) {
+        // st.w is the dictionary history after rdic(), not a spare output
+        // buffer; reusing it as output would alias reads with writes and
+        // corrupt the result. Only take the fast path without a dictionary.
+        if (!dic && st.w.length == st.u) {
           bufs.push(buf = st.w);
           ol += st.u;
         }
diff --git a/tests/simple_cases_test.ts b/tests/simple_cases_test.ts
index 04fcb55..bb86380 100644
--- a/tests/simple_cases_test.ts
+++ b/tests/simple_cases_test.ts
@@ -109,4 +109,16 @@ Deno.test("Decompression of Lorem ipsum text using a dictionary", () => {
     let shouldBeLoremIpsum = new TextDecoder().decode(output);
     assertEquals(expected, shouldBeLoremIpsum);
   }
-});
\ No newline at end of file
+});
+Deno.test("Dictionary decompression when dictionary length equals output size", () => {
+  // 200-byte raw dictionary; payload is identical to it, so it compresses to a
+  // single back-reference into the dictionary and decodes back to the dict.
+  const dictionary = new Uint8Array(200);
+  for (let i = 0; i < 200; i++) dictionary[i] = (i * 7 + 13) % 256;
+  const compressed = new Uint8Array([
+    0x28, 0xB5, 0x2F, 0xFD, 0x24, 0xC8, 0x4D, 0x00, 0x00, 0x10, 0x0D, 0x14,
+    0x01, 0x00, 0xC3, 0x25, 0x65, 0xB0, 0x43, 0xDF, 0x16, 0xB6
+  ]);
+  const decompressed = fzstd.decompress(compressed, undefined, dictionary);
+  assertEquals(dictionary, decompressed);
+});