GitLab

Both demand analysis and worker/wrapper currently don't cope with existentials, let alone GADTs:

{-# LANGUAGE GADTs #-}

module Lib where

data Box a where
  Box :: a -> Box a

data Ex a where
  Ex :: e -> a -> Ex a

data GADT a where
  GADT :: Int -> GADT Int

f :: Box Int -> Int
f (Box n) = n
{-# NOINLINE f #-}

g :: Ex Int -> Int
g (Ex _ n) = n
{-# NOINLINE g #-}

h :: GADT a -> Int
h (GADT n) = n
{-# NOINLINE h #-}

This is the Core output you get:

$wf = \ ww_sxG -> ww_sxG
f = \ w_sxA ->
      case w_sxA of { Box ww1_sxD ->
      case ww1_sxD of { I# ww3_sxG ->
      case $wf ww3_sxG of ww4_sxL { __DEFAULT -> I# ww4_sxL }
      }
      }

g = \ ds_dws ->
      case ds_dws of { Ex @ e_avJ ds1_dwB n_aud -> n_aud }

$wh = \ @ a_sxO ww_sxS ww1_sxW -> ww1_sxW
h = \ @ a_sxO w_sxP ->
      case w_sxP of { GADT ww1_sxS ww2_sxT ->
      case ww2_sxT of { I# ww4_sxW ->
      case $wh @~ <Co:1> ww4_sxW of ww5_sy1 { __DEFAULT -> I# ww5_sy1 }
      }
      }

To my surprise, the GADT in h got properly unboxed! The existential constructor in g not so much. I wonder why, GADT seems strictly more complex than Ex in its desugared form.