Possible performance regression in 3.0.10

[calmyournerves]

Hey there 👋

After upgrading from 2.2.19 (jit-mode) to 3.0.22, we saw a performance regression when compiling Tailwind CSS via esbuild/PostCSS. We have configured esbuild to compile single CSS files with an onLoad plugin that calls PostCSS and Tailwind CSS. I am not really sure if it's even intended to work like this by the creators of Tailwind CSS, but it worked great for us. Initial compile times were blazingly fast (~2s) and because of single file compilation, subsequent compiles (for development) would be very fast as well (~200ms) because we could cache and only compile what has changed.

However: After digging a little bit why it is suddenly slower, I enabled debug outputs with DEBUG=true. In the output, I saw a Setting up new context... line for every (?) compiled file. Compared to that, the output of 2.2.19 only contains it once. So I guess 2.2.19 is reusing the context and 3.0.22 isn't. After searching in the code, I found the changes in #6580 (released in 3.0.10) to be the reason for this. Because all files containing an apply directive are added to the context, it can't be reused anymore. This increases our initial compile times to roughly ~8s. Removing the the applyDirectives.size > 0 condition from the PR yields the same result as before (or even a little bit faster).

Any ideas how to tackle this? Is there a way to create a context manually and pass it to Tailwind CSS somehow?

[calmyournerves]

Should be fixed by #7524, thanks!