This repository demonstrates some issues related to compilation speed for Haskell source files containing lots of string literals. I brought this up on the Haskell Discourse recently.

There are three source files here:

Slow.hs uses the IsString instance for Builder directly.

Medium.hs uses the IsString instance for ByteString.

Fast.hs concatenates all the string literals into one big ByteString, and extracts substrings by using drop and take.

String.hs uses plain, non-overloaded String.

I'm not aware of a "native" way to measure per-file compilation times in Stack or Cabal, so I have provided the Perl script time-lines.pl. It can be used like this:

stack clean
stack build 2>&1 | ./time-lines.pl

On my machine (an AMD Ryzen™ 7 5700U running Ubuntu 20.04.3 LTS), I get:

[1 of 4] Compiling Experiment.Fast                 ( 0.62 s)
[2 of 4] Compiling Experiment.Medium               ( 0.26 s)
[3 of 4] Compiling Experiment.Slow                 ( 2.03 s)
[4 of 4] Compiling Experiment.String               ( 0.35 s)

Note: In the original version of this repo (664488615dbf30c78af34266a6f5ec8aa5cb650f), Medium was actually slower than Fast. However, now that I have changed Medium and Fast to return ByteString instead of Builder, Fast has gotten slower (0.62 seconds now versus 0.25 before) and Medium has gotten faster (0.25 seconds now versus 1.55 before). So now they are somewhat misnamed, since Fast is now slower than Medium. Don't know why that small change made such a big difference.

The stack.yaml specifies lts-18.27, which implies GHC 8.10.7 and bytestring-0.10.12.0.

I have filed two issues against the bytestring package:

• ByteString literals slow down compilation a lot
• Builder literals are noticeably slower to compile than ByteString literals