Compile-time regression between GHC-9.2 and GHC-9.4 on generically deriving TextShow instances

Summary

When porting a client project from ghc-9.2 (first to 9.6, then additionally to 9.4 to get more data), I noticed a pretty drastic compile time regression where the overall compile time of the project doubled (from roughly 8 to roughly 16 minutes). After a bit of investigation, I noticed that the slowdown is limited to a handful of modules. One of the worst affected modules had a number of derivations for TextShow instances using DerivingVia. So I'm singling out this part for now and reporting it as a bug.

After talking to @mpickering, I also did some additional experiments in how far the text version being used affects the overall outcome.

Steps to reproduce

Here is a program that can be used to reproduce: it depends on base and text-show. I think even a smaller program with just a few of the datatypes already demonstrates the problem, but the time differences are less clearly observable, so I've opted for the larger sample here:

{-# LANGUAGE DeriveGeneric #-}
{-# LANGUAGE DerivingStrategies #-}
{-# LANGUAGE DerivingVia #-}
module Slow2 where

import GHC.Int
import qualified GHC.Generics as GHC

import TextShow
import TextShow.Generic

data SM =
    CS !Int32 !Int32
  | NS !Int32
  deriving stock GHC.Generic
  deriving TextShow via FromGeneric SM

data EM =
    CE (Int, Int) (Int, Int)
  | NE (Int, Int)
  deriving stock GHC.Generic
  deriving TextShow via FromGeneric EM

data AA = K | D | M
  deriving stock GHC.Generic
  deriving TextShow via FromGeneric AA

data AWA =
    SI Int AA
  | SP Int [AWA]
  deriving stock GHC.Generic
  deriving TextShow via FromGeneric AWA

data EEP =
  EEP
    { eepP :: Int
    , eepO :: Int
    , eepB :: EM
    , eepAWA :: [[AWA]]
    }
  deriving stock GHC.Generic
  deriving TextShow via FromGeneric EEP

Expected behavior

I would hope to see no drastic regression in compile times. As it is, here is a table of compile times I have observed on my machine, split across various GHC versions and whether text-1 or text-2 is being used:

GHC version	Text version	Compile time
9.2.7	text-1.2.5.0	17.2s
9.4.5	text-1.2.5.0	35.1s
9.6.2	text-1.2.5.0	38.7s
9.2.7	text-2.0.2	17.6s
9.4.5	text-2.0.2	45.7s
9.6.2	text-2.0.2	50.7s

As can be seen, there is already a significant performance hit by just switching GHC versions, but it seems to be an even bigger hit if text-2 is used (which originally I upgraded at the same time as the GHC version, making the perceived hit even bigger).

Environment

All on Linux, GHC versions and text versions as specified above.