review (S) for 6915557: assert(_gvn.type(l)->higher_equal(type), "must constrain OSR typestate") with debug build

Tom Rodriguez Thomas.Rodriguez at Sun.COM
Wed Feb 24 22:04:26 PST 2010

On Feb 24, 2010, at 6:13 PM, Vladimir Kozlov wrote:

> Tom,
> So it is safe to skip the type check at this place because
> in an other place (in ciTypeFlow) we already have such check?
> Your new comment says that we don't do type check because it will fail anyway
> but it does not explain why it is safe to skip it at that place.

In our current system it's illegal to compile an OSR with a live jsr and it's ciTypeFlows responsibility to figure that out.  See ciTypeFlow::JsrSet::apply_control.  So any address in the OSR entry state must be invalid.  A precise version of method liveness or a more complicated version of ciTypeFlow, would be able to prove that but since ours can't prove this we have to ignore them ourselves.  This is something we've always been exposed to but it happened in the past that the copy of the OSR block that we chose to use as the entry had those locals as bottom instead of containing an address.  Because of cloning caused by jsr/ret that are hundreds of copies of that block, some which say that local is bottom and the others all say it's some constant.  If we chose a one with a constant address then we will hit this assert.


> Thanks,
> Vladimir
> Tom Rodriguez wrote:
>> 6915557: assert(_gvn.type(l)->higher_equal(type),"must constrain OSR typestate") with debug build
>> Reviewed-by:
>> The fix for 6892079 to eliminate asserts about address types in OSR
>> was insufficient because sometimes method liveness may consider locals
>> to live that actually aren't because of the conservativeness of its
>> analysis.  I think the fix is simply to not check address types.  If
>> the OSR entry point has any live address we will fail this test but
>> the resonsibility for making sure that we don't actually have a live
>> jsr during OSR entry is managed by ciTypeFlow itself so it should be
>> safe to simply skip the check.  Tested with failing test case.
>> src/share/vm/opto/parse1.cpp

More information about the hotspot-compiler-dev mailing list