Build time (measured by executing time make -j 16 from a clean state):

For a compiler without #333 and #332 (as it was before)

real	3m13.943s
user	9m41.184s
sys	2m56.065s

For a compiler with #332

real	9m46.765s
user	20m29.501s
sys	3m12.723s

For this PR

real	6m11.362s
user	14m37.577s
sys	3m4.446s

Doesn't this suggest that O3 is now too aggressive?

We're currently looking at this a bit more in depth but it mostly seems to indicate that something might be wrong with the inlining depth.

I've benchmarked the time stage2 was taking for different values of inlining_depth:

> 2:
real	1m16.495s
> 4: 
real 1min29s
> 5: 
real	2m3.569s
> 6:
real	3m36.364s
> 10:
real	4m23.641s

Or in a graph:

With an inlining depth of 6 for -O2 it looks like:

• simplify_binary_primitive.ml takes 90s to compile
• flambda_kind.ml takes 14s to compile
• numeric_types.ml takes 20s to compile