If a gem is located in a different drive than the Gemfile, standalone
mode will fail to generate the `bundler/setup` script, failing with an
error like
```
ArgumentError: different prefix: "C:/" and "D:/a/rubygems/rubygems/bundler/tmp/2/bundled_app/bundle/bundler"
C:/hostedtoolcache/windows/Ruby/3.0.5/x64/lib/ruby/3.0.0/pathname.rb:528:in `relative_path_from'
D:/a/rubygems/rubygems/bundler/tmp/2/gems/system/gems/bundler-2.4.20/lib/bundler/installer/standalone.rb:58:in `gem_path'
D:/a/rubygems/rubygems/bundler/tmp/2/gems/system/gems/bundler-2.4.20/lib/bundler/installer/standalone.rb:33:in `block (2 levels) in paths'
D:/a/rubygems/rubygems/bundler/tmp/2/gems/system/gems/bundler-2.4.20/lib/bundler/installer/standalone.rb:32:in `map'
D:/a/rubygems/rubygems/bundler/tmp/2/gems/system/gems/bundler-2.4.20/lib/bundler/installer/standalone.rb:32:in `block in paths'
```
I'm fixing this by falling back to using a full path in this case.
This was caught by a failing spec, so I'm not adding new specs.
https://github.com/rubygems/rubygems/commit/3cb9b9ab7a
For some reason, it's been failing only on YJIT, but apparently it's
reproducible on the interpreter as well. So it's not related to YJIT.
rbs marked rbs tests on Ruby master as allow_failures
https://github.com/ruby/rbs/pull/1536, so it's known to not work on Ruby
master.
We should revert this once we fix the flaky test failure on Ruby master.
Rails uses IPAddr#include? to evaluate what it should use as the
client's remote ip by filtering potential ips against a trusted list
of internal ips. In a _very_ minimal app, #include? was showing up in
a profile as ~1% of request time.
The issue is that #include? was converting itself and the other value
passed in to ranges of IPAddr. This mean as a worst case (where other is
a non-IPAddr, like a String) then there would be 5 IPAddr instances
created (other -> IPAddr, and two each for the conversions to ranges).
However, wrapping the begin and end values as IPAddr is not needed
because they are necessarily fixed addresses already.
This patch extracts the logic for getting the begin_addr and end_addr
from the #to_range method so that they can be used in #include? without
having to instantiate so many IPAddr.
Benchmark:
```ruby
net1 = IPAddr.new("192.168.2.0/24")
net2 = IPAddr.new("192.168.2.100")
net3 = IPAddr.new("192.168.3.0")
net4 = IPAddr.new("192.168.2.0/16")
Benchmark.ips do |x|
x.report("/24 includes address") { net1.include? net2 }
x.report("/24 not includes address") { net1.include? net3 }
x.report("/16 includes /24") { net4.include? net1 }
x.report("/24 not includes /16") { net1.include? net4 }
x.compare!
end
```
Before:
```
Comparison:
/24 not includes /16: 175041.3 i/s
/24 not includes address: 164933.2 i/s - 1.06x (± 0.00) slower
/16 includes /24: 163881.9 i/s - 1.07x (± 0.00) slower
/24 includes address: 163558.4 i/s - 1.07x (± 0.00) slower
```
After:
```
Comparison:
/24 not includes /16: 2588364.9 i/s
/24 not includes address: 1474650.7 i/s - 1.76x (± 0.00) slower
/16 includes /24: 1461351.0 i/s - 1.77x (± 0.00) slower
/24 includes address: 1425463.5 i/s - 1.82x (± 0.00) slower
```
* [YARP] Reject numbered parameters in block parameters
* [YARP] Do not allow BEGIN except the toplevel
---------
Co-authored-by: Haldun Bayhantopcu <haldun@github.com>
fix memory leak in vm_method
This introduces a unified reference_count to clarify who is referencing a method.
This also allows us to treat the refinement method as the def owner since it counts itself as a reference
Co-authored-by: Peter Zhu <peter@peterzhu.ca>
NODE_ARGS, NODE_ARYPTN, NODE_FNDPTN manage memory of their
structure by imemo tmpbuf Object.
However rb_ast_struct has reference to NODE. Then these
memory can be freed directly when rb_ast_struct is freed.
This commit reduces parser's dependency on CRuby functions.
