RFR: 8310130: C2: assert(false) failed: scalar_input is neither phi nor a matchin reduction [v2]

[Emanuel Peter]

On Thu, 22 Jun 2023 12:58:23 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> Removed a spurious assert before optimization bailout.
>> 
>> I assumed that the "scalar input" of a Reduction node must always be either a Phi or another Reduction node. But that is incorrect, partial vectorization can lead to a reduction node chain where we have vector reductions and scalar reductions at the same time. In those cases, we cannot move the UnorderedReductions out of the loop, so a optimization bailout is appropriate.
>> 
>> I assessed the other asserts in `PhaseIdealLoop::move_unordered_reduction_out_of_loop`, and I think they are all justified. However, one assert would have lead to a `continue` in production, which would not break out of the nested loop correctly. I changed it to a `return`, so that would be a bailout from the optimization. This assert should not be triggered because in `SuperWord::mark_reductions` we forbid that a reduction node has any uses inside the loop except for the successor node in the reduction chain.
>> 
>> I have one regression test delivered by the fuzzer, and one that I constructed myself after understanding the issue.
>> Testing up to tier6 and stress testing. **Running, passing except for some IR rules I had to fix, rerunning...**
>
> Emanuel Peter has updated the pull request incrementally with one additional commit since the last revision:
> 
>   removed unnecessary flags from test

Just saw a x86 (32bit) IR rule failure for `compiler.loopopts.superword.TestUnorderedReductionPartialVectorization.test1`. Will investigate.

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

PR Comment: https://git.openjdk.org/jdk/pull/14494#issuecomment-1606977614