Arity raising for record types

[konradbausch]

No description provided.

[b-studios]

Hey, thanks this looks great.

1. do you have some insights on how this affects performance?
2. could you delete the metals.sbt files from git?

[konradbausch]

Thanks!

It seems to affect performance positively. Our current benchmarks that use records are not running faster, since as far as I saw, they get inlined or constant-folded. But they also aren't slower. I came up with some examples that show how performance increases when the changes actually persist through compilation. The record-passing example program might seem a bit constructed, so I also added a matrix determinant example, hoping it's a bit more realistic to an actual program.

Here are the results for doing 10 warmup runs and then 50 normal runs:

record_passing

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
main	199.1 ± 4.4	189.8	210.4	9.18 ± 0.33
konradbausch/arity-raising	21.7 ± 0.6	21.0	24.0	1.00

matrix_determinant

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
main	378.3 ± 8.8	362.5	408.3	14.34 ± 0.53
konradbausch/arity-raising	26.4 ± 0.8	25.2	28.1	1.00

It seems to only have improvements on the LLVM backend, and I'm not sure why yet. I haven't tried to find examples with negative performance effects yet and will look into that as well.

I've also removed the metals.sbt files.