RFR: 8330158: C2: Loop strip mining uses ABS with min int

[Aleksey Shipilev]

On Wed, 17 Apr 2024 15:49:59 GMT, Roland Westrelin <roland at openjdk.org> wrote:

>> This fixes 3 calls to ABS with a min int argument. I think all of them
>> are harmless:
>> 
>> - in `PhaseIdealLoop::exact_limit()`, I removed the call to ABS. The
>>   check is for a stride of 1 or -1.
>>   
>> - in `OuterStripMinedLoopNode::adjust_strip_mined_loop()`, for the
>>   computation of `scaled_iters_long`, the stride is passed to `ABS()`
>>   and then implicitly casted to long. I now cast the stride to long
>>   before `ABS()`. For a min int stride, `LoopStripMiningIter * stride`
>>   overflows the int range for all values of `LoopStripMiningIter`
>>   except 0 or 1. Those values are handled early on in that method. So
>>   for a min in stride:
>>   ```
>>   (jlong)scaled_iters != scaled_iters_long
>>   ```
>>   is always true and the method returns early.
>>   
>> - in `OuterStripMinedLoopNode::adjust_strip_mined_loop()`, the
>>   computation of `short_scaled_iters` also calls `ABS()` with the
>>   stride as argument. But the result of that computation is only used
>>   if the test for:
>>   ```
>>   (jlong)scaled_iters != scaled_iters_long
>>   ```
>>   doesn't cause an early return of the method. I reordered statements
>>   so the `ABS()` calls happens after that test which will cause an early
>>   return if the stride is min int.
>
> Thanks for reviewing this.
> 
>> Have you tried running tests with #18751 applied?
> 
> I only ran the particular test that you mentioned in the bug.

@rwestrel, if you could integrate this, we can then go forward with #18751. Thanks!

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

PR Comment: https://git.openjdk.org/jdk/pull/18813#issuecomment-2068855426