Alexis King authored Apr 06, 2020 and Herbert Valerio Riedel committed Apr 13, 2020

This adds INLINABLE pragmas to most exported combinators, which enables
cross-module specialization of the Stream constraint (which can in turn
enable further optimizations). This improves performance of these
combinators in scenarios where GHC chooses not to inline them, since
they may still be specialized instead.

This change is primarily in response to a performance regression
discovered by the GHC performance test suite when running haddock (since
haddock uses parsec). The full discussion is available here:

    ghc/ghc!3041

The gist is that, without these pragmas, performance relies too heavily
on inlining heuristics working out in our favor, and subtle changes in
the optimizer can cause regressions.

The GHC performance tests suggest this patch reliably reduces runtime of
haddock on base by 7–9% and allocation by 3–5%. Pretty good for doing
something so simple!