GitLab

A module loaded by a plugin poisons the module cache, so we never load an orphan RULE or instance even if we legitimately should do so. The way to test this is a bit convoluted, but it goes something like this:

plugins07.hs

module Main where

import Plugins07a
import RuleDefiningPlugin

{-# NOINLINE x #-}
x = "foo"

main = putStrLn (show x)

Plugins07a.hs

{-# OPTIONS_GHC -fplugin RuleDefiningPlugin #-}
module Plugins07a where

RuleDefiningPlugin.hs, in ANOTHER PACKAGE (otherwise, EPT rules don't apply)

module RuleDefiningPlugin where

import GhcPlugins

{-# RULES "unsound" forall x. show x = "SHOWED" #-}

plugin :: Plugin
plugin = defaultPlugin

I'll commit the full test.

Here's what happens:

• Building Plugins07a results in loadPluginInterface on RuleDefiningPlugin. We load the ModIface but only add its types to the environment because of the "Care with plugin imports" special case.
• Building plugins07.hs results in a normal source level import for RuleDefiningPlugin, but ModIface is already in the cache so we don't load anything. RULE is not loaded, disaster!

Admittedly, actually triggering this bug requires a convoluted chain of events. But really this problem arose because the "Care with plugin imports" fix is just completely nonsense. Here's what we should do:

• We should apply the same fix from #2182 (closed) on orphan instances to orphan rules too. This way, we can safely load RULEs into the EPS without accidentally bringing them into scope when they shouldn't be.
• Loading an interface should unconditionally suck in the instances and rules.

The result is more correct and simpler, so it seems worth fixing.