Skip to content

HTTPS clone URL

Subversion checkout URL

You can clone with
or
.
Download ZIP
branch: master
Commits on Jul 2, 2015
  1. @headius
  2. @headius
  3. @headius
  4. @headius

    Tweaks for @k77ch7 fix for #3035.

    headius authored
  5. @headius
  6. @headius

    Minor cleanup.

    headius authored
  7. @headius
  8. @headius
  9. @headius
Commits on Jun 30, 2015
  1. @headius

    Kill unused method invoker logic and related classes.

    headius authored
    We have largely committed to JRuby 9k being Java 7+, and to
    simplify the process of binding jitted method bodies we are using
    MethodHandle exclusively. Because of this, and because MHs work
    anywhere reflection works, we are removing the bytecode-generate
    CompiledMethod and the Reflected*Method.
    
    There's also a bit of peripheral cleanup for code that is defunct
    with these classes removed.
  2. @headius
  3. @headius

    Always use variableNames for sizing growable scopes. Fixes #3089.

    headius authored
    variableNames is updated when IRScope gets set into StaticScope,
    and is used for other sizing and definition logic, so make it the
    sole place we get variable counts from when sizing scopes.
  4. @headius
  5. @headius

    Fix hanging in modified FIFO read/write spec, tagged in 04db9d8.

    headius authored
    In MRI, rb_read_internal always just does a blind read on the
    descriptor, since kernel-level interruption of that blocking call
    is safe. On JDK, interruption of an NIO channel read generally
    closes that channel, because there's too much state between Java
    and the native side to ensure has been cleaned up correctly. Our
    normal behavior when dealing with NIO channels is then to always
    use select before blocking reads, so we can safely interrupt when
    there's no data available. However, this is not appropriate for
    native descriptors, like the FIFO in this case, because all ENXIO
    channels appear to be selectable (so we attempt it) and the
    additional select does not match MRI semantics here.
    
    My change is to modify our version of rb_read_internal to never
    select when using a native descriptor. ENXIO should not damage the
    descriptor when interrupted, and we should still be able to
    interrupt it properly.
    
    Note also that we generally use ENXIO native channels for the
    following cases, so only these cases should see any changed
    behavior:
    
    * stdio
    * some pipes
    * most files
Commits on Jun 15, 2015
  1. @headius

    Also tag this spawn spec.

    headius authored
  2. @headius
  3. @headius
  4. @headius
  5. @headius

    Add GzipReader#readbyte. Fixes #3041

    headius authored
    Conflicts:
    	core/src/main/java/org/jruby/ext/zlib/JZlibRubyGzipReader.java
  6. @headius
Commits on Jun 5, 2015
  1. @headius

    Merge pull request #3023 from bbrowning/slow_suites

    headius authored
    Fix the slow test indexes so test:slow_suites runs tests again
Commits on Jun 3, 2015
  1. @headius

    Repurpose jit.maxSize to add a size limit for jitting methods.

    headius authored
    jruby/jruby#3016 was caused by too-large methods entering the JIT
    pipeline and consuming too much memory, mostly within ASM's data
    structures. In this case one particularly large method was the
    Ragel-generated "advance" method in @whitequark's parser library,
    which amounts to over 17k IR instructions.
    
    The default instruction count max of 2000 here is based on running
    'gem install rails', 'rails new ...' and 'rake test:mri:jit' and
    monitoring methods which exceed the limit; during this process,
    all but a handful of methods are under 2000 instructions. We
    believe this limit will allow important methods to still JIT while
    avoiding extremely large outliers causing JIT pipeline or memory
    pressure.
    
    Fixes #3016.
  2. @headius
  3. @headius
  4. @headius
Commits on Jun 2, 2015
  1. @headius

    Prefer ENV for proxy settings, because:

    headius authored
    * It is a new feature in 2.0+ stdlib, so our old logic using
      ENV_JAVA as default is now invalid.
    * http.nonProxyHosts defaults to a non-empty value, making it
      impossible to fall back on ENV.
    
    Fixes failing MRI test on master.
  2. @headius
  3. @headius
  4. @headius
  5. @headius

    This condition does not seem necessary anymore. Fixes #3008.

    headius authored
    I believe this condition was added by @tdeuhr before he adjusted
    the lookup logic to work like MRI. Without it, all tests and specs
    appear to pass as before, and the error from #3008 goes away.
  6. @headius

    Add support for http.nonProxyHosts and rework ENV_JAVA support.

    headius authored
    Fixes jruby/jruby#2983.
    
    Note that Ruby 1.9 has no support for environment-driven proxy
    settings, so this change is only available in 2.0 mode on JRuby
    1.7.
  7. @headius

    Add support for http.nonProxyHosts and rework ENV_JAVA support.

    headius authored
    Fixes jruby/jruby#2983.
    
    Note that Ruby 1.9 has no support for environment-driven proxy
    settings, so this change is only available in 2.0 mode on JRuby
    1.7.
Commits on May 28, 2015
  1. @headius
  2. @headius
  3. @headius
Something went wrong with that request. Please try again.