The previous implementation did the following (after skipping any ID3
headers at the start of the stream):
1. Skip forward byte-by-byte looking for a sync word (0xFFF)
2. Assume this indicates the start of an ADTS frame and read the size
a) If frameSize <= 6 immediately return false
3. Skip forward by frameSize and expect to find another ADTS sync word
(with no further scanning).
b) If we find one, great! Loop from step 2.
a) If we don't find one then assume the **last** sync word we found
wasn't actually one, so loop from step 1 starting one extra byte
into the stream. This means we're looking for a sync word we would
have skipped over in step 3.
The asymmetry here comes from the different handling of frameSize <= 6
(immediately return false) and frameSize being 'wrong because it doesn't
lead to another sync word' (scan the file again from the beginning for
alternative sync words).
With this change both these cases are handled symmetrically (always scan
for alternative sync words). Step 2a) becomes the same as 3b): Loop back
to the beginning of the stream with an incremented offset and scan for
another sync word.
#minor-release
PiperOrigin-RevId: 397285756
73aece6356