I tend to disagree with this. I've optimized build times on number of occasions on different projects, each being in the venue of multi-million LoC, and none of them benefited greatly from employing precompiled headers, which essentially should help alleviate exact this gap.
Running the build with -ftime-trace almost always has shown in my experiments that the bottlenecks are either in exploding number of template instatiations or more often than not in linking time.
Compiler authors can often make compilation faster, but they have to work at it and think carefully through the problem. It has to be a focus not an afterthought.
Years ago the GNAT Ada compiler authors intentionally avoided implementing precompiled headers. They instead focused on very fast lexical analysis (using a small handwritten lexer optimized for lower case letters because that was the usual case). By careful optimization of a small part of the compiler they made the recompilation of headers faster than loading a precompiled header!! More info: https://dwheeler.com/essays/make-it-simple-dewar.html
Maybe the Rust team needs to add timers to their test suite. This kind of big regression in compilation time should have been caught by the test suite.
I assure you it's not an afterthought. We track performance changes every week, going as far as reverting highly desired features to avoid regressions. The only reason we accept (small) regressions is for correctness fixes. This regression was only partly exercised by our perf test suite. That's how we grow it: see bad behaviour in the wild, add to the suite.
The source of rust compilation slowness usually boil down to excercising a part of the language that has O(n²) behaviour, like match arm evaluation, trait bound evaluation, or huge expansion due to monomorphization or macro expansion, and the resulting huge linking times from that, or the bottleneck that proc-macros introduce (they get compiled before they can get executed, only relevant on fresh compiles).
The regression was (partly) fixed by adding caching: https://github.com/rust-lang/rust/pull/90423/files and if I recall correctly was introduced by a feature change that fixed a bunch of incorrectly rejected deeply nested trait bounds.
In general, Ada has a pretty good compilation model. It has a formalized version of C++'s physical design with spec/body separation. It's also defined grammatically to required the context clause with dependencies first before everything else in a compilation unit (subunits or library units). `separate` also allows you to separately compile those things which might change a lot.
Running the build with -ftime-trace almost always has shown in my experiments that the bottlenecks are either in exploding number of template instatiations or more often than not in linking time.