Commits on Mar 5, 2015

1. Fix bug in LVA + simplify it …

   * LVA was attempting to analyze closures in the context of their
     parent scope's LVA context. The code was general enough that if
     we knew more information about specific call sites, it would have
     done better.
   
     However, in reality, given that we don't know too much about
     call sites, (a) all that code was extra complexity (b) it was
     really doing anything in terms of not being conservative (because
     all calls are being treated as dataflow barriers right now)
     (c) in the only case where it would not have been conservative,
     it was actually making an incorrect assumption that the closure
     will execute at the call site where it is defined and hence
     incorrectly marking a variable dead when it wasn't.
   
   * The fix was to simply eliminate the analysis of closures in the
     context of the parent's scope. We already had support for
     independent LVA of closures which takes care of scenario (c) above
     by marking all dirtied vars live on "entry" (which in the case of
     LVA, which is a backwards pass, is the control flow exit).
   
     This now fixes the failing snippet (see below) and should also
     eliminate the remaining test:mri:jit failures.
   [subbu@earth ir] jruby -Xjit.threshold=0 -e "z = 5; x = Hash.new {|h,kk| z = kk }; z=0; x[22]; p z"
   22
   
   Conflicts:
   	core/src/main/java/org/jruby/ir/dataflow/analyses/LiveVariableNode.java
   	core/src/main/java/org/jruby/ir/passes/LiveVariableAnalysis.java

   

   subbuss committed Mar 5, 2015
   Copy the full SHA

   f48c871 View commit details

   Browse the repository at this point in the history

2. FIC build is complete when linearized bbs are ready

   

   subbuss committed Mar 5, 2015
   Copy the full SHA

   699c833 View commit details

   Browse the repository at this point in the history