Goals for the first usable iteration of -Zself-profile are:

• Make the compiler track query invocations and other important function calls (e.g. LLVM related)
  • This means tracking the query/function name (no query keys, arguments yet)
• Reduce the overhead of tracking and profile generation as to not be prohibitive
  • This means emitting events in an optimized binary format
• Write a post-processing tool that generates an aggregated report from the raw event data
  • The aggregated report is a table with one line per query/function and columns for
    • total time spent in the query (in milliseconds)
    • time spent in the query as percentage of total compile time
    • number of query invocations
    • percentage of in-memory cache hits
    • percentage of incremental cache hits
    • total time spent (milliseconds) in loading query results from incremental cache
    • total time spent (milliseconds) blocked on concurrent query invocations
• Re-enable self-profiling on perf.rlo, which includes
  • running the postprocessing tool to generate the report for each test run
  • adding a new comparison view that compares the test runs of a single benchmark and shows changes per query. This view is reachable by clicking on a benchmark in the regular comparison view (i.e. one can "zoom" into a given benchmark)
• Document how self-profiling works in the rustc-guide.

Non-Goals are:

• Supporting self-profiling in 32-bit compilers -- this makes it easier to rely on things like memory mapped files
• Tracking individual query keys/function arguments

Work packages resulting from this set of goals are:

• Implement a library that takes care of reading and writing the binary event format
  • See https://github.com/rust-lang/measureme/
• Make the compiler use the library to emit profiling data efficiently
  • Initial integration at Use measureme in self profiler #59515
  • Implement "event filtering" in order to keep profiling overhead low in the common case. (Implement event filtering for self-profiler. #59915)
  • Add output directory argument to -Zself-profile (Allow to specify profiling data output directory as -Zself-profile argument. #61123)
  • Add a version header to profiler artifacts (Add versioning to the binary profile format measureme#40)
• Implement a postprocessing tool (using the library) that generates the aggregated report (Implement a summarization tool for profile traces measureme#17)
• Make perf.rlo support self-profile:
  • run benchmarks with -Zself-profile
  • run postprocessing tool
  • store aggregated reports
  • implement the detailed comparison view
  • make the regular comparison view link to detailed views
• Review and make sure that we are tracking everything we are interested in. Things to check:
  • Pre-query passes (parsing, macro expansion, name resolution, HIR lowering, ...)
  • LLVM optimization passes
  • Metadata loading/decoding
  • Trait selection (removed this from the MVP for now)
• Document how self-profiling works in the rustc-guide.
• Polishing iteration
  • Write high-level crate docs for measureme (implemented in Write some crate level documentation measureme#68)
  • Detailed view should show "percentage of total time" column (detailed-view: Percentage of total time column missing rustc-perf#523)
  • Show total sum line in table for the entire crate (detailed-view: Add a "sum total" line at the top of the table. rustc-perf#525)
  • Make sorting more visible/accessible in the results table (detailed views: Make it more visible that columns can be sorted. rustc-perf#526)
  • It's unclear what the "invocations" and "cache misses" columns in the detailed view are exactly. (detailed (comparison) view: The cache misses column should be removed, invocations should be renamed rustc-perf#529)
  • Resolve a bug in the sum for incr. loading time column (detailed view: Possible bug in "totals" line for incr. comp. cache loading rustc-perf#527)
  • Clean up self-time computation in summarize (Event recording & summarize need cleanup with respect to self-time vs incr-load-time vs blocked-time vs total-time. measureme#75)

Possible Problems that might arise are:

• Profiling overhead keeps being too high - then we need to think about doing separate self-profile runs on perf.rlo

cc @wesleywiser @Mark-Simulacrum