RFR: 8283094: Add Ideal transformation: x + (con - y) -> (x - y) + con

[Quan Anh Mai]

On Mon, 21 Mar 2022 12:43:48 GMT, Roland Westrelin <roland at openjdk.org> wrote:

>> I think I have a better idea for this. During idealisation, we will transform every subtraction into an addition with the negation of the second operand, pushing the negation down the computation graph. This will simplify numerous rules we have with subtractions. Addition presents nice commutative and associative properties so the graph transformation during idealisation with be much easier. Before matching, we may have another strength reduction phase which will push the negations up the computation graph, transform the computations to be as compact as possible.
>> 
>> This problem seems a lot like the problem we face in #7700 when eagerly strength reducing operations hurt our ability to perform other transformations by complicating the elements inside the graph. I believe this solution would simplify our arithmetic transformations and allow better numeric calculations. Similar to what has been discussed in #7395 , a more general solution would be much more valuable than trying to match every possible combinations.
>> 
>> Thank you very much.
>
>> I think I have a better idea for this. During idealisation, we will transform every subtraction into an addition with the negation of the second operand, pushing the negation down the computation graph.
> 
> One issue with such an approach is that it's then required to update all code that pattern match an expression with SubX node to now expect an (AddX (NegX ..) ..). PhaseIdealLoop::is_scaled_iv_plus_offset() for instance. And maybe there are not that many instances of code that require updating, but it's still needed to proceed carefully and double check, possibly update some code, write tests that verify the change does what's intended etc. So what appears relatively simple is going to be more work than expected and the risk associated with the change is higher than one might think. I doubt that's worth it.

@rwestrel  Thanks a lot for your information, it seems that the idea would turn out to be problematic with little applicable benefits then.

-------------

PR: https://git.openjdk.java.net/jdk/pull/7795