RFR(M): 8210215: C2 should optimize trichotomy calculations
tobias.hartmann at oracle.com
Tue Oct 16 12:18:04 UTC 2018
thanks for the review!
I agree that optimizing pointers might confuse later optimization phases. I'm not sure either about
floats but to be on the safe side, I've excluded both floats and pointers with the comment you
suggested. We can later always adapt and re-enable the optimization for floats if required.
For the record, here's the webrev of the changes that I've just pushed:
On 15.10.2018 19:13, John Rose wrote:
> Reviewed. I *think* you could actually allow floats, because your optimization
> doesn't use all properties of trichotomy. Specifically, you don't utilize the
> "excluded middle" property: If !(x<y) and !(x>y), you don't conclude x==y,
> which is what NaNs would break. Likewise, if !(x>=y) you don't conclude
> x<y, and if !(x<=y) you don't conclude x>y. (Um, am I right about all this?)
> Excluded middles are deduced from the negations of related conditions,
> but you don't actually work with negations; everything is normalized to
> a positive condition, and then "merged" (logical-and-ed) to produce a
> further positive condition. This is why I think NaNs are probably OK.
> If you do exclude floats, please add a comment like:
> // Floats don't exactly obey trichotomy. To be on the safe side, don't transform their tests.
> Of course, if you know a case where they actually break the logic,
> document that, instead. I have a feeling there is no such case.
> Pointers don’t exactly obey trichotomy either; IIRC CmpP only produces NE and EQ states.
> I think we use GT to encode NE, since we don’t have a proper NE (which is unfortunate).
> If you see two pointers where !(p>q) and !(p==q) you can't conclude !(p<q),
> because you are actually in a paradoxical state (not reached). Well, I guess
> that means it would be OK to conclude (p<q) but you'll confuse later phases
> of the optimizer.
> Bottom line: Double-check the logic on pointers and floats, and do the right thing.
> Either way, I'm very happy with what you have made here. Thanks
> for polishing the test to use non-constant test values.
> — John
> On Oct 9, 2018, at 10:29 AM, Tobias Hartmann <tobias.hartmann at oracle.com> wrote:
>> I had an off thread discussion with John and he suggested the following changes in addition:
>> - Use randomization for integer values in the test
>> - Add some asserts and comments to improve readability
>> - Don't apply optimization to floating point compares because they don't obey trichotomy laws (for
>> example, NaN always compares to false)
>> - Handle the case where the merge result is BoolTest::illegal (just for robustness, I haven't seen
>> this ever during testing but it might show up when overflow tests mix with "normal" other tests)
>> - While testing, I found that we also have to add a call to add_users_to_worklist(iff1) to make sure
>> that the dead if is removed (otherwise we hit failures like JDK-8075922).
>> Incremental webrev:
>> Full webrev:
>> Best regards,
More information about the hotspot-compiler-dev