Eliminate 4 of 8 syscalls when requiring file by absolute path
Don't open file twice when specified by absolute path.
require '/a.rb' (i.e. via an absolute path), ruby generates this sequence of syscalls:
open /a.rb fstat64 /a.rb close /a.rb open /a.rb fstat64 /a.rb fstat64 /a.rb read /a.rb close /a.rb
It is apparent that the only inherently necessary members of this sequence are:
open /a.rb fstat64 /a.rb read /a.rb close /a.rb
(the fstat64 isn't obviously necessary, but it does serve a purpose and probably shouldn't be removed).
The first open/fstat64/close is used to check whether the file is loadable. This is important when scanning the
$LOAD_PATH, since it is used to determine when a file has been found. However, when we've already unambiguously identified a file before invoking
require, this serves no inherent purpose, since we can move whatever work is happening as a result of that
fstat64 into the second open/close sequence.
This change bypasses the first open/fstat64/close in the case of an absolute path to
require. It also removes one of the doubled-up
fstat64 calls later in the sequence. As a result, the number of syscalls to require a file changes:
- From 8 to 4 when specified by absolute path;
- From 5+3n to 4+3n otherwise (where n is the number of
In future work, it would be possible to re-use the file descriptor opened while searching the
$LOAD_PATH without the close/open sequence, but this would cause some ugly layering issues.
We intend to use this in conjunction with something like https://github.com/shopify/bootscale, which pre-resolves required features to absolute paths before calling
require. This change reduces our total number of filesystem accesses by 13% during application boot.
Various notes and rationale at http://notes.burke.libbey.me/ruby-require-optimization