How to avoid construction of dead dictionaries

Motivation

I'm writing this at Simon's direction here:

https://mail.haskell.org/pipermail/glasgow-haskell-users/2021-August/027007.html

We're using Conal Elliott's ConCat framework to implement Deep Learning. (This is a project in production automation for finding anomalies in sensor data.)

ConCat includes this definition:

(<+) :: Con a => (Con b => r) -> (a |- b) -> r
r <+ Entail (Sub Dict) = r

The right-hand argument of <+ leads to a dictionary construction that is a proof of a certain property, but the dictionary itself ends up being dead, like so:

      case $w$dOpCon_r2kGJ ...
      of
      { (# ww1_s27L3 #) -> ... }
           ^^^^^^^^^
           never used

Yet, ghc (8.10.4) never elides this code. (I'm naively assuming because of the unboxed tuple, but actually have no clue.)

You can try this out here:

https://github.com/mikesperber/concat/tree/dead-dictionaries

Do this:

stack build neural-networks-concat:lib

This generates oodles of output, where you can search for OpCon.

Proposal

I don't have a concrete proposal. But is there any chance of convincing ghc of erasing the dictionary construction?