You've seen how load-linked/store-conditional and compare-and-swap let hardware build atomic primitives one word at a time. But what if you want to atomically update ten words — say, moving a node between two linked lists? With CAS, you fake it with fine-grained locks, hazard pointers, or a monstrous multi-CAS software protocol. Hardware Transactional Memory (HTM) lets you just say: "execute these instructions atomically, or don't execute them at all."

The mechanism reuses infrastructure you already have: the cache coherence protocol and the store buffer. When software issues XBEGIN, the core enters transactional mode. Every cache line the transaction reads gets tagged in the R-set; every line it writes goes into the W-set and is held in the L1 cache in a modified-but-not-yet-visible state. Stores don't drain to L2. Loads pull normally but mark the tag.

The coherence protocol becomes the conflict detector. If another core sends a snoop that would invalidate a line in your R-set (someone wrote what you read) or read a line in your W-set (someone read what you're about to publish), the transaction aborts. Your speculative writes are discarded — because they were never written back — and the register file is rolled back from a checkpoint taken at XBEGIN. On XEND, the W-set flips atomically from "speculative" to "modified," instantly visible to the coherence fabric.

Intel's TSX (Transactional Synchronization Extensions) is the canonical real-world example. It shipped in Haswell (2013), got recalled via microcode after the TSX-NI bug, returned in Broadwell, and has been repeatedly disabled since due to side-channel vulnerabilities (TAA, most notably). IBM's POWER8 and z-Series mainframes have been running HTM in production databases (DB2) for over a decade — that's where it earns its keep.

The fundamental limit: your R-set and W-set must fit in L1 cache. A transaction touching more lines than L1 associativity supports will be evicted, and eviction of a tagged line = automatic abort. Rule of thumb: on a 32 KB, 8-way L1 with 64-byte lines, you get ~500 lines of transactional footprint — if nothing conflicts on set indexing. Real workloads hit 50-200 lines reliably. Anything bigger needs a fallback path.

Which is why every HTM library ships with a lock-based fallback: try the transaction 3-5 times, and if it keeps aborting (capacity, conflict, or interrupt), grab a real mutex. HTM is a fast path, not a replacement.