make Internal.Utilities.Text.Lexing.Position a struct

[vasily-kirichenko]

Heap allocation before:

After:

[dsyme]

Hi @vasily-kirichenko

This is potentially a good change. A couple of things

• Although allocations will reduce, it's possible that the lexer and parser become slower (because more time is spent copying structs). Could you write a perf test to repeatedly parse some input? I don't think it matters too much which input it is.
• Ideally this change would via Microsoft/visualfsharp --> fsharp/fsharp --> fsharp/FSharp.Compiler.Service to keep the codebases maximally aligned. Would you be able to submit the change to Microsoft/visualfsharp and assess its impact on perf of a compiler bootstrap?

Thanks
Don

[vasily-kirichenko]

Hi, Don.

How exactly I can run parsing on a file? If I use FSharpChecker.ParseFileInProject, it caches results and all subsequent calls returns immediately.

[dsyme]

I guess you could generate a sequence of textually different files?

[vasily-kirichenko]

I've tested it on all *.fs files in src/fsharp dir https://github.com/vasily-kirichenko/FSharp.Compiler.Service/blob/parser-perf/tests/ConsoleApplication1/Program.fs

Results: before - 6.7 s, after - 7.03 s, which is roughly 5% digression :(

Memory traffic (all objects created during the parsing), before:

after:

Now Tuple<Position, Position> takes almost 2x more memory. Maybe it's a good idea to replace it with a custom struct?

[vasily-kirichenko]

I replaced a couple of top allocated Tuples with custom structs and make a couple of functions to be top-level. After that parsing time reduced and equals 6.26 s (6% better than "before" result). Number of allocated objects reduced by 45% (!), allocated memory - by 9%. I tried to make TokenTup to be a struct, but it increased parsing time by 10%.

[dungpa]

@dsyme Have you had a second look at this PR? It seems like a good enhancement to make.

[dsyme]

I replaced a couple of top allocated Tuple s with custom structs and make a couple of functions to be top-level. After that parsing time reduced and equals 6.26 s (6% better than "before" result). Number of allocated objects reduced by 45% (!), allocated memory - by 9%.

@vasily-kirichenko Is this what's currently in the PR? Thx.

[dsyme]

(I'll reopen since there does seem to be value here, once we confirm the tradeoffs)

[vasily-kirichenko]

@dsyme yes.

[dsyme]

@vasily-kirichenko Would it be possible for you to submit this to Microsoft\visualfsharp as well please? I don't mind accepting it here first but ideally we should get it submitted there too.

[vasily-kirichenko]

@dsyme done dotnet/fsharp#664