`dunai`: `morphGS` is probably inefficient about dunai

[turion]

The branch may serve as a starting point for further future optimizations, of which arrM is only one.

[turion]

Further bisecting of all the possible inlines suggests that inlining arr is also a huge performance win.

[ivanperez-keera]

Just a note: Issue #375 addressed the introduction of a benchmark. It will require discussion before making changes, but it may be worth seeing if those benchmarks, or how they are used, should be improved.

[ivanperez-keera]

@turion thanks a lot for this very detailed report. I know it's a lot to ask on top of what you've already done, but would you be able to put together a small benchmark that shows the difference in terms of performance with one and the other?

[turion]

Yes, I managed to find a benchmark using ListT that reproduces the issue, I believe. It's much more visible there because many MSFs are launched.

Reproduce with cabal run spaceleak --enable-profiling -- +RTS -hc -l-agu -i0.01 && eventlog2html spaceleak.eventlog on branch https://github.com/turion/dunai/tree/dev_spaceleak.

Library on branch develop, with benchmark added

89d4125 Adding spaceleak

Replaced arrM definition with direct implementation

110e026 Write out arrM

• Huge performance win (time per tick is reduced a lot)
• Peak memory usage reduced from 28 to 21 MB
• Space leak still visible

Heavy inlining and strictification in many places (shotgun optimization)

11ba331 Inline many functions

• Again huge performance win
• Peak memory usage further reduced
• Space leak still visible
• Unclear yet which of the many optimizations had an effect, further research necessary

[turion]

My current strategy to identify the minimal set of necessary changes is:

1. Bisect on the profiling result until a commit is identified that improves performance
2. Cherry-pick that commit onto a branch improvements
3. Rebase the whole branch with all inlinings onto improvements
4. Repeat until the performance improving commit is already on improvements

[turion]

Trying the benchmarks briefly seems to show that these changes here have a significant impact on the benchmarks. I think it would be very helpful if we could plot benchmarks (+ error bars) for several commits to compare them.