Commits on Jul 4, 2024

1. [NO-TICKET] Optimize CodeProvenance#record_loaded_files to avoid allo… …

   …cations
   
   **What does this PR do?**
   
   This PR replaces a `Hash#find` with a `Hash#any?` in the Profiler
   `CodeProvenance` collector.
   
   This seemingly innocuous change actually gets rid of a bunch of memory
   allocations.
   
   It turns out that when `Hash#find` is being used, Ruby allocates an
   array to contain the `[key, value]` pair that's passed into the block,
   for each entry in the hash (on top of a few more objects -- but this is
   the biggest offender unless the hash is empty/almost empty).
   
   The VM actually has an optimization for a lot of operations, including
   `Hash#any?` that avoids allocating any extra memory, but because
   `Hash#find` is actually inherited from `Enumerable`, this optimization
   does not kick in.
   
   **Motivation:**
   
   Eliminate unneeded memory allocations.
   
   **Additional Notes:**
   
   Here's a very simple reproducer that shows the issue:
   
   ```
   puts RUBY_DESCRIPTION
   
   def allocated_now; GC.stat(:total_allocated_objects); end
   
   hash = {a: 1, b: 2, c: 3}
   v = nil
   
   before_find = allocated_now
   10.times { _, v = hash.find { |k, _| k == :c } }
   puts "Found value #{v}, allocated #{allocated_now - before_find}"
   
   before_any = allocated_now
   10.times { v = nil; hash.any? { |k, val| (v = val) if k == :c } }
   
   puts "Found value #{v}, allocated #{allocated_now - before_any}"
   ```
   
   and here's how it looks on the latest Ruby version:
   
   ```
   ruby 3.4.0preview1 (2024-05-16 master 9d69619623) [x86_64-linux]
   Found value 3, allocated 64
   Found value 3, allocated 1
   ```
   
   (Note that there's a few more allocations going on with `Hash#find`,
   not only the arrays for the pairs, and we get rid of them ALL!)
   
   **How to test the change?**
   
   This operation already has test coverage. I considered adding some
   counting of allocated objects, but we've had quite a bunch of
   flakiness in some of the profiler specs when counting objects, so
   I've decided to not add any performance-specific tests for this.

   

   ivoanjo committed Jul 4, 2024
   Configuration menu

   • View commit details
   • Copy full SHA for f0be554
   • Browse repository at this point

   Copy the full SHA

   f0be554 View commit details

   Browse the repository at this point in the history