profile with xcode instruments

[xusd320]

Problem

Mako is a both I/O intensive and CPU intensive program, directly reading the source codes and writing outputs with std::fs will cause the main compile blocking.

Analysis

TODO: Use tools like Xcode Instruments to track the I/O syscalls.

Solution

Let I/O and compiler run in individual thread pool.

[xusd320]

Performance analysis with instruments

File Activity

• Cold build (without os kernel fs cache):

  • fs read:
    those long timing fs reads and fs stats are been invoked by mako::load::read_content and nodejs_resolver.
    

    • in mako::load::read_content, there is only a call of std::fs::read_to_end，maybe no better implements. The crate memmap2 is a choice, but also will leads to some risks.
    
    • a lots of invalid readlink and stat calls invoked by nodejs_resolver, why ?
    

  • fs write:
    fs write is fast enough when building examples/with-antd, maybe no need to optimize. Not test on heavy project.
    

• Hot build:
  Still spends a lot of time on fs reads
  

Timing Profiler:

• Too much Regex::new, is being solved by @PeachScript
  

System Trace:

• Threads:
  Threads are not well load balanced,  a lot of blocking occurs
  
  
  What's those tokio workers doing for?

• Sys calls:

  • Those are spending on rayon and tokio threads.
    

  • Those are spending on mako::load::read_content and nodejs_resolver. About nodejs_resolver probleam, test it again after switch to oxc-resolver.
    

  • Mutex
    

[xusd320]

tree shaking 和 generate 在主线程中串行执行，在构建 example/with-antd 时，这两个阶段各占总耗时的 1/4