<p><span class="h-card" translate="no"><a href="https://mastodon.social/@whitequark" class="u-url mention">@<span>whitequark</span></a></span> nice.</p><p>back when working on Souper I had a longstanding hilarious interaction with some LLVM folks where I'd ask them for an IR-level cost model and they'd tell me one and it wouldn't work at all. we never really arrived at a very good answer.</p><p>Minotaur lowers IR to assembly and uses llvm-mca to predict cost in cycles or uOps, this works fine</p>
