perf: use mimalloc for bindings_napi

[sapphi-red]

This PR contains:

• bugfix
• feature
• refactor
• documentation
• other

Are tests included?

• yes (bugfixes and features will not be merged without tests)
• no

Breaking Changes?

• yes (breaking changes will not be merged unless absolutely necessary)
• no

List any relevant issue numbers:

Description

This PR swaps the memory allocator with mimalloc.
This allocator seems to be common in napi-rs world (swc, @napi-rs/fast-escape, parcel, @resvg/resvg-js).
Some projects uses jemallocator on some OS/architectures and ideally we should check which allocator performs better on each of them. But I'll leave this investigation for now.

Running the benchmark in #5182 (actually I used the esbuild repo instead), I got ~200ms / ~3% improvement on my laptop.

Benchmark results

rollup 3.29.4

# BUILD: 6079ms, 934 MB / 944 MB
## initialize: 1ms, 3.55 kB / 9.85 MB
## generate module graph: 2891ms, 765 MB / 775 MB
- plugin 0 (stdin) - resolveId: 16ms, -3.18 MB / 773 MB
- plugin 0 (stdin) - load: 7ms, 883 kB / 774 MB
generate ast: 632ms, 159 MB / 774 MB
analyze ast: 1382ms, 829 MB / 774 MB
## sort and bind modules: 253ms, 38.4 MB / 813 MB
## mark included statements: 2934ms, 131 MB / 944 MB
treeshaking pass 1: 1013ms, 102 MB / 917 MB
treeshaking pass 2: 457ms, 24 MB / 941 MB
treeshaking pass 3: 184ms, -5.31 MB / 936 MB
treeshaking pass 4: 178ms, 6.55 MB / 942 MB
treeshaking pass 5: 198ms, -115 kB / 942 MB
treeshaking pass 6: 169ms, 14.5 MB / 957 MB
treeshaking pass 7: 164ms, -5.48 MB / 951 MB
treeshaking pass 8: 152ms, 3.51 MB / 955 MB
treeshaking pass 9: 137ms, 1.66 MB / 956 MB
treeshaking pass 10: 137ms, 2.53 MB / 959 MB
treeshaking pass 11: 139ms, -15.3 MB / 944 MB
# GENERATE: 553ms, 107 MB / 1.05 GB
## initialize render: 0ms, 3.46 kB / 946 MB
## generate chunks: 47ms, 4.11 MB / 950 MB
optimize chunks: 1ms, 394 kB / 950 MB
## render chunks: 487ms, 83.1 MB / 1.03 GB
## transform chunks: 19ms, 19.7 MB / 1.05 GB
## generate bundle: 0ms, 11.7 kB / 1.05 GB
# WRITE: 31ms, 19.7 MB / 1.07 GB

rollup 4.0.2

# BUILD: 6656ms, 917 MB / 926 MB
## initialize: 0ms, 3.55 kB / 9.06 MB
## generate module graph: 3437ms, 747 MB / 756 MB
- plugin 0 (stdin) - resolveId: 13ms, 3.18 MB / 755 MB
- plugin 0 (stdin) - load: 8ms, 882 kB / 756 MB
generate ast: 1177ms, 137 MB / 756 MB
analyze ast: 1419ms, 721 MB / 756 MB
## sort and bind modules: 245ms, 38.3 MB / 794 MB
## mark included statements: 2973ms, 132 MB / 926 MB
treeshaking pass 1: 1123ms, 103 MB / 899 MB
treeshaking pass 2: 464ms, 24.5 MB / 924 MB
treeshaking pass 3: 175ms, -5.22 MB / 918 MB
treeshaking pass 4: 173ms, 6.72 MB / 925 MB
treeshaking pass 5: 196ms, -284 kB / 925 MB
treeshaking pass 6: 161ms, 14.4 MB / 939 MB
treeshaking pass 7: 150ms, -5.39 MB / 934 MB
treeshaking pass 8: 136ms, 3.57 MB / 937 MB
treeshaking pass 9: 119ms, 1.7 MB / 939 MB
treeshaking pass 10: 124ms, 2.52 MB / 942 MB
treeshaking pass 11: 147ms, -15.3 MB / 926 MB
# GENERATE: 573ms, 107 MB / 1.04 GB
## initialize render: 0ms, 3.46 kB / 928 MB
## generate chunks: 50ms, 3.55 MB / 932 MB
optimize chunks: 1ms, 428 kB / 933 MB
## render chunks: 503ms, 82.5 MB / 1.01 GB
## transform chunks: 19ms, 20.8 MB / 1.04 GB
## generate bundle: 0ms, 1.74 kB / 1.04 GB
# WRITE: 33ms, 19.5 MB / 1.05 GB

rollup 4.1.0

# BUILD: 6597ms, 847 MB / 855 MB
## initialize: 0ms, 3.55 kB / 8.51 MB
## generate module graph: 3451ms, 667 MB / 676 MB
- plugin 0 (stdin) - resolveId: 16ms, -6.69 MB / 677 MB
- plugin 0 (stdin) - load: 8ms, 882 kB / 677 MB
generate ast: 1208ms, 168 MB / 677 MB
analyze ast: 1404ms, 650 MB / 677 MB
## sort and bind modules: 246ms, 37.3 MB / 713 MB
## mark included statements: 2900ms, 142 MB / 855 MB
treeshaking pass 1: 1086ms, 100 MB / 815 MB
treeshaking pass 2: 474ms, 22.1 MB / 837 MB
treeshaking pass 3: 183ms, 10.3 MB / 847 MB
treeshaking pass 4: 171ms, 6.99 MB / 854 MB
treeshaking pass 5: 194ms, -688 kB / 853 MB
treeshaking pass 6: 160ms, -1.58 MB / 852 MB
treeshaking pass 7: 140ms, -5.54 MB / 846 MB
treeshaking pass 8: 137ms, 3.62 MB / 850 MB
treeshaking pass 9: 116ms, 1.5 MB / 851 MB
treeshaking pass 10: 119ms, 2.55 MB / 854 MB
treeshaking pass 11: 110ms, 1.15 MB / 855 MB
# GENERATE: 569ms, 101 MB / 958 MB
## initialize render: 0ms, 3.46 kB / 857 MB
## generate chunks: 49ms, 1.97 MB / 859 MB
optimize chunks: 1ms, 380 kB / 860 MB
## render chunks: 503ms, 72.7 MB / 932 MB
## transform chunks: 16ms, 26.2 MB / 958 MB
## generate bundle: 0ms, 1.74 kB / 958 MB
# WRITE: 33ms, 17.5 MB / 975 MB

using mimalloc (~200ms / ~3% improvement)

# BUILD: 6345ms, 853 MB / 862 MB
## initialize: 0ms, 3.55 kB / 9.11 MB
## generate module graph: 3204ms, 671 MB / 680 MB
- plugin 0 (stdin) - resolveId: 13ms, 1.27 MB / 679 MB
- plugin 0 (stdin) - load: 9ms, 882 kB / 680 MB
generate ast: 996ms, 163 MB / 680 MB
analyze ast: 1392ms, 722 MB / 680 MB
## sort and bind modules: 220ms, 37.7 MB / 718 MB
## mark included statements: 2920ms, 144 MB / 862 MB
treeshaking pass 1: 1039ms, 101 MB / 820 MB
treeshaking pass 2: 480ms, 23.6 MB / 844 MB
treeshaking pass 3: 189ms, 10.3 MB / 854 MB
treeshaking pass 4: 171ms, -9.57 MB / 845 MB
treeshaking pass 5: 209ms, -27.3 kB / 844 MB
treeshaking pass 6: 161ms, 14.2 MB / 859 MB
treeshaking pass 7: 143ms, -5.51 MB / 853 MB
treeshaking pass 8: 146ms, 3.66 MB / 857 MB
treeshaking pass 9: 120ms, 1.59 MB / 858 MB
treeshaking pass 10: 130ms, 2.51 MB / 861 MB
treeshaking pass 11: 124ms, 1.18 MB / 862 MB
# GENERATE: 578ms, 113 MB / 961 MB
## initialize render: 0ms, 3.46 kB / 848 MB
## generate chunks: 51ms, 3.72 MB / 852 MB
optimize chunks: 1ms, 481 kB / 853 MB
## render chunks: 510ms, 82.8 MB / 935 MB
## transform chunks: 17ms, 26.2 MB / 961 MB
## generate bundle: 0ms, 1.74 kB / 961 MB
# WRITE: 35ms, 15.4 MB / 976 MB

Specs of my laptop

• CPU: Intel Core-i7 1360P
• Memory: DDR5-4800 32GB
• OS: Windows 11

This change increases the binary size by 0.07MB (3.25MB -> 3.32MB).

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
rollup	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Oct 14, 2023 3:17pm

[codecov]

Codecov Report

Merging #5201 (bda838f) into master (cb144b2) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #5201   +/-   ##
=======================================
  Coverage   98.82%   98.82%           
=======================================
  Files         231      231           
  Lines        8849     8849           
  Branches     2314     2314           
=======================================
  Hits         8745     8745           
  Misses         43       43           
  Partials       61       61           

[lukastaegert]

Very nice, these are the kind of improvements that I hoped people with Rust experience would dig up.

[github-actions]

This PR has been released as part of rollup@4.1.1. You can test it via npm install rollup.

[jerome-benoit]

@sapphi-red: the linux arm64 arch exclusion is intentional?

[sapphi-red]

Yes, it is. The examples I shown in the OP all excludes linux+musl+arm64. I don't know the reason, but as I cannot easily check the actual behavior, I went with a safer way.

[jerome-benoit]

That issue microsoft/mimalloc#556 seems to say that since version 2.0.5, it works on linux arm64. So if the mimalloc-rust bindings is against that version or superior, it should also work.

[sapphi-red]

The rust bindings uses 1.7.9: https://github.com/LemonHX/mimalloc-rust#:~:text=mimalloc%201.7.9%20stable
So that seems to be the reason.