RFR: 8323220: Reassociate loop invariants involved in Cmps and Add/Subs

[Emanuel Peter]

On Thu, 11 Jan 2024 19:59:39 GMT, Joshua Cao <duke at openjdk.org> wrote:

>> // inv1 == (x + inv2) => ( inv1 - inv2 ) == x
>> // inv1 == (x - inv2) => ( inv1 + inv2 ) == x
>> // inv1 == (inv2 - x) => (-inv1 + inv2 ) == x
>> 
>> 
>> For example,
>> 
>> 
>> fn(inv1, inv2)
>>   while(...)
>>       x = foobar()
>>       if inv1 == x + inv2
>>            blackhole()
>> 
>> 
>> We can transform this into
>> 
>> 
>> fn(inv1, inv2)
>>   t = inv1 - inv2
>>   while(...)
>>       x = foobar()
>>       if t == x
>>            blackhole()
>> 
>> 
>> Here is an example: https://github.com/openjdk/jdk/blob/b78896b9aafcb15f453eaed6e154a5461581407b/src/java.base/share/classes/java/lang/invoke/LambdaFormEditor.java#L910. LHS `1` and RHS `pos` are both loop invariant
>> 
>> Passes tier1 locally on Linux machine. Passes GHA on my fork.
>
> test/hotspot/jtreg/compiler/loopopts/InvariantCodeMotionReassociateCmp.java line 191:
> 
>> 189:     @Arguments({Argument.NUMBER_42, Argument.NUMBER_42})
>> 190:     @IR(failOn = {IRNode.SUB_I})
>> 191:     public void leDontReassociate(int inv1, int inv2) {
> 
> I added DontReassociate tests for `le`, `gt`, and `ge`. For `lt`, C2 generates a second `SUB_I` as part of other transformations.
> 
> IR matching for ADD/SUB is pretty hard in general. They commonly are created as part of other transformations. Any suggestions on how I can test this better is appreciated.

You could always use a simple regex with the linenumber I guess. But that is a bit nasty too.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/17375#discussion_r1453737924