Skip to content

Profiling ​

Compiling oxlint in release mode with debug information ​

For profiling, you will need to compile the oxlint binary in release mode with debug information enabled. You can do that by passing --profile release-with-debug to cargo build:

bash
cargo build --profile release-with-debug --bin oxlint

After building, the binary is located at ./target/release-with-debug/oxlint. This is the binary that should be used for profiling.

CPU - Samply ​

Samply is a command line CPU profiler which uses the Firefox profiler as its UI. Works on macOS and Linux.

To use Samply with oxlint, run samply record followed by the oxlint command and arguments:

bash
samply record ./target/release-with-debug/oxlint .

To improve the profiling experience, you might consider some of the following options:

  • oxlint: --silent will suppress diagnostics output and keep the profile more focused.
  • oxlint: --threads 1 will run the linter single threaded, which is slower, but makes it easier to analyze the profile for single-threaded performance.
  • samply record: --rate <number> will sample the profile at a higher rate. The default is 1000Hz (1ms), but increasing this will provide more detailed information at the cost of a larger profile file.

For example, running oxlint single-threaded with a 0.1ms sample rate:

bash
samply record --rate 10000 ./target/release-with-debug/oxlint --silent --threads 1 .

CPU - Mac Xcode Instruments ​

cargo instruments is the tool of choice to bridge Mac Xcode instruments.

The following instruction replicates the procedure of cargo instruments.

First, install Xcode Instruments command-line tools:

bash
xcode-select --install

Then, if you haven't already, ensure that the oxlint binary is compiled.

Under the hood, cargo instruments invokes the xcrun xctrace command, which is equivalent to

bash
xcrun xctrace record --template 'Time Profile' --output . --launch -- /path/to/oxc/target/release-with-debug/oxlint

Running the command above produces the following output

Starting recording with the Time Profiler template. Launching process: oxlint.
Ctrl-C to stop the recording
Target app exited, ending recording...
Recording completed. Saving output file...
Output file saved as: Launch_oxlint_2023-09-03_4.41.45 PM_EB179B85.trace

Open the trace file open Launch_oxlint_2023-09-03_4.41.45\ PM_EB179B85.trace.

To see a top down trace:

  1. On the top panel, click CPUs
  2. On the left input box, click x then select Time Profiler
  3. At the bottom panel, click "Call Tree", turn on "Invert Call Tree" and turn off separate by thread.

For memory and disk operations, use --template 'Allocations' and --template 'File Activity'.

For more detailed CPU profiling, such as L1/L2 cache misses, cycle and instruction counts, and branch prediction info, you need to use a custom "CPU Counters" template:

  1. Open Instruments and select the "CPU Counters" template.
  2. In the "CPU Counters" settings:
    1. Turn on the "High Frequency Sampling" option.
    2. Below the "High Frequency Sampling" option, click the plus icon and select an event type. Some suggested event types:
      • Cycles - for getting a rough idea of how many CPU cycles are spent in each function.
      • Instructions - for getting a rough idea of how many CPU instructions are executed in each function and how many cycles that takes
      • L1D_CACHE_MISS_LD - count of L1 cache misses from loading data from memory
  3. Once you have enabled the events you are interested in, save the template in "File > Save as Template ..." and give it a name.
  4. Now you can use this with xctrace by passing the template name to the --template option: xcrun xctrace record --template 'My Custom CPU Counters' --output . --launch -- /path/to/oxc/target/release-with-debug/oxlint

Heap Allocation ​

Try dhat.

Released under the MIT License.