(0001647)
geoffclare (manager)
2013-06-08 09:12
|
It seems odd that C11 would have different thread-safety requirements
for mbrlen, mbrtowc, and wcrtomb with a null state argument than for
mblen, mbtowc, and wctomb. We should query this with the C committee,
as it may well be unintentional. |
(0001648)
dalias (reporter)
2013-06-08 12:14
|
I think there's a very good reason for the discrepancy: the restartable versions can store a partially-decoded character in the mbstate_t object, so even for state-independent encodings, there is state which would need to be protected against data races. The non-restartable versions, on the other hand, are pure except in the case of state-dependent encodings, which are mostly a relic of the past and which were never supported on most POSIX systems, since these encodings are mostly incompatible with POSIX filesystem semantics. Only implementations supporting such encodings (which might not even exist - can anyone confirm?) would incur the burden of avoiding data races. Note that these functions give applications access to information on whether the locale's encoding is state-dependent, so a portable application could use the restartable interfaces when the locale is state-dependent, and the non-restartable ones otherwise.
As to the motivation behind my request for this change, I have spent a good deal of time investigating the performance bottlenecks in character-at-a-time multibyte processing, and it turns out that there is a fundamental bottleneck in the restartable interfaces due to their interface requirements for handling the ps argument and partially-decoded characters. For applications which don't need partial-character processing capability, I believe it would make sense to encourage a transition to the non-restartable interfaces, but of course this is problematic if the non-restartable interfaces are not thread-safe. In my experiments, I found the non-restartable interfaces capable of reaching roughly a 50% performance advantage over the restartable ones; this difference would of course become even more extreme if the core decoding algorithms were further optimized. |