Manuel M T Chakravarty
I noted this on D4177 and discussed the effect with Alan. Indeed there is quite a sizeable regression in compilation time but thankfully this is not because GHC itself is slower. Rather, it simply requires more work to compile. I did a set of nofib runs with and without the first TTG patch and found that compiler allocations remained essentially unchanged.
A 15% regression in the compilation time of GHC is indeed hard to stomach but Alan had said that much of this will likely disappear in the future. If this is the case then a temporary regression is in my opinion acceptable.
Hmm, on what grounds does he think that this is going to disappear and how likely is likely? This doesn’t sound convincing TBH.
As Alan has pointed out elsewhere, currently the TTG involves some unnecessarily large constraints which very likely inflate typechecking time. This is almost certainly why we are seeing such increases in compilation time of the compiler (and, as importantly, no changes in the performance characteristics of the resulting compiler). Given how consistent the story seems to be, a reduction after the constraints are simplified sounds very likely. Cheers, - Ben