Remove TemplateHaskell as a Global Default Extension

[mdimjasevic]

TemplateHaskell as a language extension can lead to unnecessary recompilations. To aid GHC in figuring out what needs recompilation, this PR removes TemplateHaskell as a global default extension and moves it to modules that do need it. Hopefully this reduces the compilation time by a noticeable amount of time.

The one-line change that triggered the rest is the removal of TemplateHaskell from package-defaults.yaml.

Checklist

• The PR Title explains the impact of the change.
• The PR description provides context as to why the change should occur and what the code contributes to that effect. This could also be a link to a JIRA ticket or a Github issue, if there is one.
• changelog.d contains the following bits of information (details):
  • A file with the changelog entry in one or more suitable sub-sections. The sub-sections are marked by directories inside changelog.d.

[fisx]

Did you take the time before and after? Either way, as long as it doesn't slow down compilation I'm in favor. :)

[mdimjasevic]

Did you take the time before and after? Either way, as long as it doesn't slow down compilation I'm in favor. :)

I didn't. At first and without too much thinking, I thought I'd see some time savings in the CI runs for this PR. However, I think it's the same time as before because everything is compiled from scratch. Now I am expecting time savings while developing as that's when we edit already compiled modules and recompile them.

@isovector brought this up in one of our pairing sessions. I don't have a good reference for this, though I tried looking it up. To some extent these are relevant:

• https://well-typed.com/blog/2022/04/hls-performance/
• https://www.parsonsmatt.org/2021/07/12/template_haskell_performance_tips.html

[mdimjasevic]

And this cannot slow down the compilation as here we're making the reach of TemplateHaskell strictly smaller, suggesting GHC knows more about what needs to be recompiled.

[mdimjasevic]

I'll proceed with merging. In the unlikely case of us having a proof this slows down the compilation, it can easily be reverted.

[fisx]

git co template-haskell-per-module
make clean
time make
git revert --no-commit a776512b0035eabf0f941ea31aa0e4d00e40fa45
make clean
time make

=>

real	8m33.984s
real	8m39.574s

i don't think that's significant, so all good!

[isovector]

The performance improvements here are going to be in incremental builds. GHC maintains interface files which hash their exports; subsequent files only need to get rebuilt if one of their imports gets changed. TemplateHaskell messes with this machinery, where it doesn't check which imports are used by splices; thus, if any dependency of a module with TH changes, the TH module will rebuild, and invalidate the build artifacts of everything downstream.