From d3168cfe3fccb4a245d99ec4069a5fc8e99060e8 Mon Sep 17 00:00:00 2001 From: BattermanZ Date: Mon, 16 Nov 2026 23:00:11 +0100 Subject: [PATCH] A cold just check and just test builds in half the time The debug profile optimised every dependency, macros and build scripts included. Only the crates the tests run hard are optimised now: photographs and WebP, inflate/deflate, the PDF reader, the HTML parser, SQLite, Argon2, hashing and case folding. From an empty cache the cargo steps of just check and just test go from 28.3 to 15.6 minutes; the warm test run is 3.7% slower. Closes #623 Co-Authored-By: Claude Opus 5.5 Claude-Session: https://claude.ai/code/session_013rPDLgbCESE7tSxLeK86m8 --- Cargo.toml | 80 ++++++++++++++++++++++++++++++++++++++++++++++-------- 1 file changed, 69 insertions(+), 11 deletions(-) diff --git a/Cargo.toml b/Cargo.toml index 502cc79..aacf564 100644 --- a/Cargo.toml +++ b/Cargo.toml @@ -227,18 +227,76 @@ test = false name = "translations_corpus" test = false -# Dependencies are compiled optimised even in a debug build; Kamosu's own crate -# is not. Re-encoding a camera-original photograph runs through `image` and -# `webp`, and unoptimised those take twenty-two seconds for one 7 MB picture — -# past the test client's read timeout, which is why the Photograph corpus check -# could not pass on any machine. Optimising only the dependencies keeps our own -# code fully debuggable and costs one slower build of libraries that then cache. +# A debug build optimises only the dependencies the tests run hard, and builds +# every other dependency and Kamosu's own crate unoptimised (#623). # -# They also carry no debug info. Kamosu's own crate keeps all of its, so a panic -# in our code still names the file and line; what goes is the ability to step +# Those few are optimised because unoptimised they are too slow to test with. +# Re-encoding a camera-original photograph through `image` and `webp` took +# twenty-two seconds for one 7 MB picture, past the test client's read timeout, +# so the Photograph corpus check could not pass on any machine. The PDF reader +# ran into its ten-second limit before its memory cap on the PDF built to swell +# it. Unoptimised, the HTML parser took thirty seconds over the page nested far +# past any real one, where optimised it takes four. WebP and resizing are much of +# a Sample Library build, and SQLite, Argon2 and case folding run on almost +# every test. +# +# Everything else stays unoptimised because optimising it cost more than it +# saved. With every dependency optimised, the builds in `just check` and `just +# test` took 24 minutes from an empty cache on four cores; with these alone +# they take 11, and the warm test run is under 4% slower. Typst alone was over +# five minutes of processor time, and the old `"*" = { opt-level = 2 }` +# optimised the macros and build scripts too. A dependency newly found slow in +# a test joins the list by name, with the crates it hands its work to. +# +# Dependencies carry no debug info. Kamosu's own crate keeps all of its, so a +# panic in our code still names the file and line; what goes is stepping # through a dependency's internals in a debugger, which nothing here does. The # saving is in linking: every test binary statically links all 635 crates, and # their debug info was the bulk of what the linker had to copy. -[profile.dev.package."*"] -opt-level = 2 -debug = 0 +[profile.dev.package] +"*" = { debug = 0 } +# Photographs: decoding, resizing, colour, and WebP out. +image = { opt-level = 2 } +image-webp = { opt-level = 2 } +png = { opt-level = 2 } +zune-jpeg = { opt-level = 2 } +zune-core = { opt-level = 2 } +fdeflate = { opt-level = 2 } +miniz_oxide = { opt-level = 2 } +crc32fast = { opt-level = 2 } +adler2 = { opt-level = 2 } +simd-adler32 = { opt-level = 2 } +moxcms = { opt-level = 2 } +webp = { opt-level = 2 } +libwebp-sys = { opt-level = 2 } +# Inflating and deflating: the PDF reader's streams, Paprika's gzip files, +# Bundles and Backups. `flate2` hands this work to `zlib-rs`. +flate2 = { opt-level = 2 } +zlib-rs = { opt-level = 2 } +# The PDF reader. +pdf-extract = { opt-level = 2 } +lopdf = { opt-level = 2 } +weezl = { opt-level = 2 } +encoding_rs = { opt-level = 2 } +cff-parser = { opt-level = 2 } +postscript = { opt-level = 2 } +adobe-cmap-parser = { opt-level = 2 } +type1-encoding-parser = { opt-level = 2 } +# Reading a web page: the importer's parse, and the article text read from a +# page with no recipe data, down a page nested past any real one. +scraper = { opt-level = 2 } +html5ever = { opt-level = 2 } +markup5ever = { opt-level = 2 } +tendril = { opt-level = 2 } +string_cache = { opt-level = 2 } +selectors = { opt-level = 2 } +cssparser = { opt-level = 2 } +ego-tree = { opt-level = 2 } +# Passwords, the database, fingerprints and case folding. +argon2 = { opt-level = 2 } +blake2 = { opt-level = 2 } +sha2 = { opt-level = 2 } +rusqlite = { opt-level = 2 } +libsqlite3-sys = { opt-level = 2 } +caseless = { opt-level = 2 } +unicode-normalization = { opt-level = 2 }