TypeError: no case for org.truffleruby.core.cast.ToFNodeGen org.truffleruby.core.numeric.RubyBignum

[rickhull]

Here is the full error text: TypeError: TruffleRuby doesn't have a case for the org.truffleruby.core.cast.ToFNodeGen node with values of type org.truffleruby.core.numeric.RubyBignum

I encountered this error as part of a testing matrix for my GitHub Action CI on a project of mine. This failure occurs for truffleruby and truffleruby-head but not for jruby or any MRI I tested. Here is the CI ouput: https://github.com/rickhull/compsci/runs/1520388175?check_suite_focus=true#logs

[eregon]

Full log (GitHub Actions logs disappear after some time):

  1) Error:
CompSci::Fit::Fit.power#test_0001_must accept power data:
TypeError: TruffleRuby doesn't have a case for the org.truffleruby.core.cast.ToFNodeGen node with values of type org.truffleruby.core.numeric.RubyBignum
	from org.truffleruby.core.cast.ToFNodeGen.executeAndSpecialize(ToFNodeGen.java:107)
	from org.truffleruby.core.cast.ToFNodeGen.executeToDouble(ToFNodeGen.java:49)
	from org.truffleruby.core.MathNodes$LogNode.function(MathNodes.java:477)
	from org.truffleruby.core.MathNodesFactory$LogNodeFactory$LogNodeGen.execute(MathNodesFactory.java:3231)
	from org.truffleruby.language.arguments.CheckArityNode.execute(CheckArityNode.java:41)
	from org.truffleruby.language.methods.ExceptionTranslatingNode.execute(ExceptionTranslatingNode.java:33)
	from org.truffleruby.language.RubyRootNode.execute(RubyRootNode.java:61)
	from org.graalvm.compiler.truffle.runtime.OptimizedCallTarget.executeRootNode(OptimizedCallTarget.java:559)
    /home/runner/work/compsci/compsci/lib/compsci/fit.rb:146:in `log'
    /home/runner/work/compsci/compsci/lib/compsci/fit.rb:146:in `block in power'
    /home/runner/work/compsci/compsci/lib/compsci/fit.rb:48:in `map'
    /home/runner/work/compsci/compsci/lib/compsci/fit.rb:48:in `sigma'
    /home/runner/work/compsci/compsci/lib/compsci/fit.rb:146:in `power'
    /home/runner/work/compsci/compsci/test/fit.rb:134:in `block (5 levels) in <top (required)>'
    /home/runner/work/compsci/compsci/test/fit.rb:132:in `each'
    /home/runner/work/compsci/compsci/test/fit.rb:132:in `block (4 levels) in <top (required)>'
    /home/runner/work/compsci/compsci/test/fit.rb:131:in `each'
    /home/runner/work/compsci/compsci/test/fit.rb:131:in `test_0001_must accept power data'

[eregon]

Looks like we need negated guards for

truffleruby/src/main/java/org/truffleruby/core/MathNodes.java

Line 471 in 8ed19c0

protected double function(Object a, NotProvided b,

There is already a specialization handling RubyBignum but it isn't taken (because the Object specialization was activated first).

Also ToFNode independently should try to convert any Object too, not just RubyDynamicObject.

@bjfish Could you do that?

[rickhull]

I am looking at the test case in my codebase. I do have a conditional guard on certain values, noting: Fit.error warning: Bignum out of Float range

Depending on platform and implementation details, it seems "legitimate" that certain Bignum's cannot be represented accurately by a float. It's possible this failure has to do with Github's platform moreso than TruffleRuby itself. I'm trying to put together a testcase that doesn't depend on my codebase.

Specifically:

  # y = ax^b
  describe "Fit.power" do
    it "must accept power data" do
      [0.01, 7.5, 500, 1000, 5000, 9999].each { |a|
        [-114, -100, -10, -0.5, -0.1, 0.1, 0.75, 10, 50, 60].each { |b|
          next if b == -114 # Fit.error warning: Bignum out of Float range
          ary = Fit.power(@xs, @xs.map { |x| a * x**b })
          expect(ary[0]).must_be_close_to a
          expect(ary[1]).must_be_close_to b
          expect(ary[2]).must_equal 1.0
        }
      }
    end
  end

Notice I skip the b == -114 case. At b == -115 this code starts generating NaNs, which is why -114 was chosen. Having removed the guard for -114, I am not seeing the Bignum warnings anymore, on my local machine. It's possible that -114 always worked locally and was only a problem for CI (Travis at the time). Not sure if this is relevant or helpful, but thought I would note.

[eregon]

@rickhull This case is a clear bug, at least in the LogNode, there is already handling for RubyBignum due to some missing check in the generic case.

This is enough to reproduce the bug:
ruby -e 'class C<Numeric; def to_f; 1.0; end; end; obj=C.new; [obj, (1<<65)].each { |e| p Math.log(e) }'

[rickhull]

Glad I could help!

[bjfish]

@rickhull This is fixed at f3b1e5c. Thanks for reporting this.