fix: do not set BOM seen if decode() errored

Author: ChALkeR

fallback/encoding.js

@@ -100,10 +100,11 @@ export class TextDecoder {
         }
       }
 
+      let seenBOM = false
       if (this.#canBOM) {
         const bom = this.#findBom(prefix ?? u)
         if (bom) {
-          if (stream) this.#canBOM = false
+          seenBOM = true
           if (prefix) {
             prefix = prefix.subarray(bom)
           } else {
@@ -117,7 +118,8 @@ export class TextDecoder {
       if (!this.#decode) this.#decode = unicodeDecoder(this.encoding, !this.fatal)
       try {
         const res = (prefix ? this.#decode(prefix) : '') + this.#decode(u) + suffix
-        if (res.length > 0 && stream) this.#canBOM = false
+        // "BOM seen" is set on the current decode call only if it did not error, in "serialize I/O queue" after decoding
+        if (stream && (seenBOM || res.length > 0)) this.#canBOM = false
         return res
       } catch (err) {
         this.#chunk = null // reset unfinished chunk on errors

tests/encoding/mistakes.test.js

@@ -713,6 +713,16 @@ describe('Common implementation mistakes', () => {
         t.assert.throws(() => fatal.decode(u(0xfd, 0xef), { stream: true }))
         t.assert.strictEqual(fatal.decode(), '')
       }
+
+      // Chrome fails
+      {
+        const d = new TextDecoder('utf-8', { fatal: true })
+        t.assert.strictEqual(d.decode(u(0xea, 0xaf, 0x8d)), '\uABCD')
+        t.assert.strictEqual(d.decode(u(0xea), { stream: true }).length, 0)
+        t.assert.throws(() => d.decode(u(0xea, 0xaf, 0x8d), { stream: true }))
+        t.assert.strictEqual(d.decode(u(0xea, 0xaf, 0x8d), { stream: true }), '\uABCD')
+        t.assert.strictEqual(d.decode(u(), { stream: true }).length, 0)
+      }
     })
 
     test('utf-8 BOM handling', (t) => {
@@ -724,12 +734,34 @@ describe('Common implementation mistakes', () => {
         t.assert.strictEqual(d.decode(u(0xef, 0xbb, 0xbf)).length, 0)
       }
 
+      // Bun fails
       {
         const d = new TextDecoder('utf-8', { fatal: true })
         t.assert.strictEqual(d.decode(u(0xef, 0xbb, 0xbf), { stream: true }).length, 0) // BOM
         t.assert.throws(() => d.decode(u(0xff), { stream: true }))
         t.assert.strictEqual(d.decode(u(0xef, 0xbb, 0xbf)).length, 1)
       }
+
+      // WebKit, FireFox, Servo, Deno got this wrong. Chrome and Node.js pass
+      {
+        const d = new TextDecoder('utf-8', { fatal: true })
+        t.assert.throws(() => d.decode(u(0x20, 0xff), { stream: true }))
+        t.assert.strictEqual(d.decode(u(0xef, 0xbb, 0xbf)).length, 0)
+        t.assert.throws(() => d.decode(u(0xff), { stream: true }))
+        t.assert.strictEqual(d.decode(u(0xef, 0xbb, 0xbf)).length, 0)
+        t.assert.throws(() => d.decode(u(0xef, 0xbb, 0xbf, 0xff), { stream: true }))
+        t.assert.strictEqual(d.decode(u(0xef, 0xbb, 0xbf)).length, 0)
+      }
+
+      // Chrome has this wrong
+      {
+        const d = new TextDecoder('utf-8', { fatal: true })
+        t.assert.strictEqual(d.decode(u(0xef), { stream: true }).length, 0)
+        t.assert.throws(() => d.decode(u(0xef, 0xbb, 0xbf), { stream: true }))
+        t.assert.strictEqual(d.decode(u(0xef, 0xbb, 0xbf), { stream: true }).length, 0)
+        t.assert.strictEqual(d.decode(u(0xef, 0xbb, 0xbf), { stream: true }).length, 1)
+        t.assert.strictEqual(d.decode(u(), { stream: true }).length, 0)
+      }
     })
 
     // https://github.com/facebook/hermes/pull/1855#issuecomment-3632349129