perf(tspath): optimize the fast path case for ToFileNameLowerCase

[camc314]

When using project services, ToFileNameLowerCase is called 10, 312, 056 times on various paths, this means that it's incredibly hot path.

Since most projects won't hav e non-ascii filenames, we can optimize for this case. This yields up to a 5x perf improvement in some scenarios.

Please feel free to close, if the perf degregation againt non-ascii file paths outweighs the perf gains for the common case

Test Case	Likelihood Category	Old (ns)	New (ns)
All lowercase ASCII short	High	45.1	8.5
All uppercase ASCII short	Medium	71.8	37.9
Mixed case ASCII short	High	62.2	41.9
Mixed case ASCII longer	High	134.0	60.9
Non-ASCII without İ (e.g., ß)	Low	152.0	166.7
Non-ASCII with İ	Very low	147.5	156.8
Non-ASCII with ı	Low	128.2	138.1
Long mixed ASCII	Low	1260.6	547.8

[jakebailey]

Can you include some benchmarks that show the difference? As in, Go benchmarks in path_test.go?

I'm not surprised by this PR, of course; strings.ToLower makes this same optimization, we just can't use it directly. (It might be worth simply using the same code as ToLower and tweaking it for the special case.)

[camc314]

Can you include some benchmarks that show the difference? As in, Go benchmarks in path_test.go?

mind clarifying exactly what your looking for?

typescript-go/internal/tspath/path_test.go

Line 545 in 4be54c5

func BenchmarkToFileNameLowerCase(b *testing.B) {

exists already, which is what i've been using to bench the perf differences. are you looking for a bench function that compares the two (old vs new impl)

[camc314]

@microsoft-github-policy-service agree

[camc314]

@microsoft-github-policy-service agree

[jakebailey]

Ah, yes. I didn't realize that's what you were doing. I think I was expecting results through https://pkg.go.dev/golang.org/x/perf/benchstat.

Comparing main and this PR with go test -run=- -bench='BenchmarkToFileNameLowerCase' -benchmem -count=10 ./internal/tspath:

goos: linux
goarch: amd64
pkg: github.com/microsoft/typescript-go/internal/tspath
cpu: Intel(R) Core(TM) i9-10900K CPU @ 3.70GHz
                                                     │   old.txt    │               new.txt               │
                                                     │    sec/op    │   sec/op     vs base                │
ToFileNameLowerCase//path/to/file.ext-20               58.495n ± 2%   9.892n ± 1%  -83.09% (p=0.000 n=10)
ToFileNameLowerCase//PATH/TO/FILE.EXT-20               101.25n ± 1%   45.84n ± 0%  -54.73% (p=0.000 n=10)
ToFileNameLowerCase//path/to/FILE.EXT-20                85.81n ± 2%   46.68n ± 1%  -45.60% (p=0.000 n=10)
ToFileNameLowerCase//user/UserName/proje...etc-20      186.45n ± 1%   97.24n ± 1%  -47.84% (p=0.000 n=10)
ToFileNameLowerCase//user/UserName/proje...etc#01-20    206.2n ± 1%   222.8n ± 1%   +8.05% (p=0.000 n=10)
ToFileNameLowerCase//user/UserName/proje...etc#02-20    202.5n ± 1%   215.2n ± 1%   +6.25% (p=0.000 n=10)
ToFileNameLowerCase//user/UserName/proje...etc#03-20    174.4n ± 1%   190.2n ± 1%   +9.06% (p=0.000 n=10)
ToFileNameLowerCase/FoO/FoO/FoO/FoO/FoO/...etc-20      1623.5n ± 1%   952.5n ± 1%  -41.33% (p=0.000 n=10)
geomean                                                 180.4n        107.5n       -40.38%

                                                     │   old.txt    │                new.txt                │
                                                     │     B/op     │    B/op     vs base                   │
ToFileNameLowerCase//path/to/file.ext-20               0.000 ± 0%     0.000 ± 0%         ~ (p=1.000 n=10) ¹
ToFileNameLowerCase//PATH/TO/FILE.EXT-20               24.00 ± 0%     24.00 ± 0%         ~ (p=1.000 n=10) ¹
ToFileNameLowerCase//path/to/FILE.EXT-20               24.00 ± 0%     24.00 ± 0%         ~ (p=1.000 n=10) ¹
ToFileNameLowerCase//user/UserName/proje...etc-20      48.00 ± 0%     96.00 ± 0%  +100.00% (p=0.000 n=10)
ToFileNameLowerCase//user/UserName/proje...etc#01-20   48.00 ± 0%     48.00 ± 0%         ~ (p=1.000 n=10) ¹
ToFileNameLowerCase//user/UserName/proje...etc#02-20   48.00 ± 0%     48.00 ± 0%         ~ (p=1.000 n=10) ¹
ToFileNameLowerCase//user/UserName/proje...etc#03-20   48.00 ± 0%     48.00 ± 0%         ~ (p=1.000 n=10) ¹
ToFileNameLowerCase/FoO/FoO/FoO/FoO/FoO/...etc-20      416.0 ± 0%     832.0 ± 0%  +100.00% (p=0.000 n=10)
geomean                                                           ²                +18.92%                ²
¹ all samples are equal
² summaries must be >0 to compute geomean

                                                     │   old.txt    │                new.txt                │
                                                     │  allocs/op   │ allocs/op   vs base                   │
ToFileNameLowerCase//path/to/file.ext-20               0.000 ± 0%     0.000 ± 0%         ~ (p=1.000 n=10) ¹
ToFileNameLowerCase//PATH/TO/FILE.EXT-20               1.000 ± 0%     1.000 ± 0%         ~ (p=1.000 n=10) ¹
ToFileNameLowerCase//path/to/FILE.EXT-20               1.000 ± 0%     1.000 ± 0%         ~ (p=1.000 n=10) ¹
ToFileNameLowerCase//user/UserName/proje...etc-20      1.000 ± 0%     2.000 ± 0%  +100.00% (p=0.000 n=10)
ToFileNameLowerCase//user/UserName/proje...etc#01-20   1.000 ± 0%     1.000 ± 0%         ~ (p=1.000 n=10) ¹
ToFileNameLowerCase//user/UserName/proje...etc#02-20   1.000 ± 0%     1.000 ± 0%         ~ (p=1.000 n=10) ¹
ToFileNameLowerCase//user/UserName/proje...etc#03-20   1.000 ± 0%     1.000 ± 0%         ~ (p=1.000 n=10) ¹
ToFileNameLowerCase/FoO/FoO/FoO/FoO/FoO/...etc-20      1.000 ± 0%     2.000 ± 0%  +100.00% (p=0.000 n=10)
geomean                                                           ²                +18.92%                ²
¹ all samples are equal
² summaries must be >0 to compute geomean

Which I think matches what you were quoting. Surprised the other cases get slower, honestly.

[camc314]

Ah, yes. I didn't realize that's what you were doing. I think I was expecting results through https://pkg.go.dev/golang.org/x/perf/benchstat.

yess, although I can write go, it's been a while and I am far from an expert, i'm also not up to date with the latest tooling, so there's definitly cases where I could be using a tool to do what i'm doing manually 😅

Surprised the other cases get slower, honestly.

the other cases will have slowed down because we now have to iterate over the string twice, this shouldn't be a problem in 99% of cases since most file paths will be ascii, and won't contain that special i character

[jakebailey]

the other cases will have slowed down because we now have to iterate over the string twice, this shouldn't be a problem in 99% of cases since most file paths will be ascii, and won't contain that special i character

Or at least for usernames with non-ASCII characters, they'll be early in the path via /home/<name> or /Users/<name> or C:/Users/<name>..

[jakebailey]

yess, although I can write go, it's been a while and I am far from an expert, i'm also not up to date with the latest tooling, so there's definitly cases where I could be using a tool to do what i'm doing manually 😅

For the record, I do:

$ go test -run=- -bench='BenchmarkToFileNameLowerCase' -benchmem -count=10 ./internal/tspath | tee old.txt
# switch branch
$ go test -run=- -bench='BenchmarkToFileNameLowerCase' -benchmem -count=10 ./internal/tspath | tee new.txt
$ benchstat old.txt new.txt

[jakebailey]

Just out of curiosity, where did you experience this being a bottleneck?

[Boshen]

Just out of curiosity, where did you experience this being a bottleneck?

We are diving into

• Performance issue with project service oxc-project/tsgolint#67

i.e. running tsgolint in large monorepos (1000+ files).

[camc314]

Just out of curiosity, where did you experience this being a bottleneck?

Yep to add to Boshen's point, i'm testing this on vscode repo (~6100) files, during which, this function is called 10 million + time.

thanks for the extra bechmarking info Jake, i'll be using that next time!

[jakebailey]

Fuzz test is clean after running for a bit.

(We don't check these in CI, it would be good if we did though they take a while and you need to select a specific test to fuzz, not just ask for all to be fuzzed.)