Linked in the Bugzilla thread is a really nice in depth investigation of the same issue with high register aliases in a similar algorithm (Huffman coding) but in an entirely different product: https://fgiesen.wordpress.com/2025/05/21/oodle-2-9-14-and-in... .

It's concerning that Intel don't seem to have been responsive to anyone with respect to this issue and it doesn't appear to have an official errata yet.

Details of the errata from a comment in the diff:

"Write both dist bytes as a single 2-byte store. This avoids the `movb %ch, [mem]` instruction pattern (store from high-byte register alias) that LLVM otherwise emits when dist arrives as a wide register. That pattern triggers the Intel Raptor Lake CPU errata, causing silent 2-byte stores that corrupt the adjacent `len` byte."

Uh ... working around this in each and every piece of software sounds like a non-starter? Intel should be on the hook to fix this.