Backtracking regexes have pathological worst-case performance when
a long line contains a large amount of whitespace not followed by
a newline, since the regex engine will attempt to match the regex
at each whitespace character, read ahead to the non-whitespace non-
newline, declare no match, and try again at the next whitespace.
E.g. try running
util.removeTrailingSpaces(new Array(1e6).join(' ') + 'a').length
which would hang V8.
Since we still run the full test suite on Node.js, this would
only no longer catch bugs which are specific to
- Browser non-draft04 GCM (and don't manifest in draft04 GCM,
and don't manifest in Node.js non-draft04 GCM)
- Browser OCB (which is not natively implemented in the browser)
- Browser V5 Keys
Many tests would run for every encryption mode, or for both V4 and V5 keys,
without there being any difference between the different test runs.
`grunt coverage` before and after this commit reports almost identical
statistics, providing some confidence that no code coverage was lost.
When not requested, we convert the streams to Uint8Arrays.
This makes the generated key safe to pass to a Worker more than once.
Partially reverts 735aa1da.
When encrypting/signing a stream, this allows you to indicate whether it's a
stream of Strings or Uint8Arrays (using message.fromText or message.fromBinary,
respectively.)
When signing text, this allows you to control whether to create a cleartext
message or a regular armored text message.
When creating a detached signature, it allows you to control whether it's "meant
for" (verifying against) a cleartext message. A cleartext message has trailing
whitespace trimmed before signing. This fixes the case of passing a detached
signature from sign() to encrypt(). Since encrypt() doesn't create a cleartext
message, the signature would be invalid if the text contained lines with
trailing whitespace.
