TX lock elision

Posted Jan 25, 2026 19:43 UTC (Sun) by ballombe (subscriber, #9523)
In reply to: TX lock elision by csamuel
Parent article: GNU C Library 2.43 released

So it is yet another hardware feature that does not pan out.

to post comments

TX lock elision

Posted Jan 25, 2026 20:23 UTC (Sun) by csamuel (✭ supporter ✭, #2624) [Link]

Sadly yes. Back in 2012 when we were getting our BlueGene/Q there was a lot of excitement about it having HTM but it never came to anything. There was this paper from IBM itself (PDF) https://dl.acm.org/doi/pdf/10.1145/2370816.2370836 which says:

> While it is widely expected that such overheads be significantly reduced in an HTM, one of the surprising findings of this performance study is that, the BG/Q HTM overhead, while much smaller than that of STM’s, is still non-trivial. The causes of HTM overheads are also very different from those of STM’s. For instance, BG/Q TM maintains speculative states in the L2 cache. This allows for transactions with a large memory footprint, the price to pay, however, is the overhead, where the L1 cache is either bypassed during a transaction or flushed upon entry to a transaction. The loss of cache locality is the dominant cause of BG/Q TM overhead

TX lock elision

Posted Jan 25, 2026 21:27 UTC (Sun) by josh (subscriber, #17465) [Link] (4 responses)

That's pretty much the history of transactional memory, both software and hardware.

https://imgur.com/a/cXCtzsr

TX lock elision

Posted Jan 26, 2026 6:22 UTC (Mon) by pthariensflame (subscriber, #169079) [Link] (1 responses)

For whatever it's worth, (software) transactional memory has found genuine sustained use in one specific ecosystem: Haskell, in which the monad-based effect control makes it a lot less error-prone to use.

TX lock elision

Posted Jan 27, 2026 1:49 UTC (Tue) by josh (subscriber, #17465) [Link]

I'm familiar (I used to program Haskell), but I also think it fits Haskell because it optimizes for compositional simplicity at the expense of performance.

TX lock elision

Posted Feb 8, 2026 17:33 UTC (Sun) by joib (subscriber, #8541) [Link] (1 responses)

Amusingly, it seems a patch queued up for the 6.20 (or 7.0 if Linus decides to bump the major version number?) kernel enables it on x86, with a mention in the commit msg stating that SUSE has been having it enabled for a long time.

https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.g...

7.0

Posted Feb 8, 2026 20:42 UTC (Sun) by corbet (editor, #1) [Link]

At the Maintainers Summit, Linus told us that the next release would be 7.0. No need to wonder about that — unless he changes his mind, of course.