Big package databases make GHC slow to start up

[ezyang]

ezyang@sabre:~$ time ghc -no-global-package-db -e "return ()" 

<interactive>:1:1:
    Failed to load interface for ‘Prelude’
    Use -v to see a list of the files searched for.

real	0m0.100s
user	0m0.080s
sys	0m0.016s
ezyang@sabre:~$ time ghc  -e "return ()" 

real	0m0.230s
user	0m0.176s
sys	0m0.052s
ezyang@sabre:~$ time ghc -package-db .cabal/store/ghc-7.10.3/package.db/ -e "return ()" 

real	0m0.310s
user	0m0.268s
sys	0m0.036s

That's a 3x difference for no discernible benefit at all! And it gets worse as the package database size increases. Maybe we should populate the plan database with exactly the packages involved in our plan.

[hvr]

Fwiw, we will also need to figure out a way to make ghc-pkg register not become a bottleneck for big nix-stores (at least for future GHC versions, if we can't help existing GHCs). I'll soon run into the situation where I'll have several thousands of packages in a huge nix-store, and there ghc-pkg will incur a significant overhead when cabal registers a new package.