RFR: 8326139: C2 SuperWord: split packs (match use/def packs, implemented, mutual independence) [v2]

[Vladimir Kozlov]

On Wed, 21 Feb 2024 08:36:11 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> Emanuel Peter has updated the pull request with a new target base due to a merge or a rebase. The pull request now contains 27 commits:
>> 
>>  - manual merge
>>  - more fixes for TestSplitPacks.java
>>  - fix some IR rules in TestSplitPacks.java
>>  - fix MulAddS2I
>>  - Fix data dependency IR rules in LoopArrayIndexComputeTest.java
>>  - fix populate index cases and test7
>>  - test7 WIP
>>  - WIP fix reductions
>>  - Update test/hotspot/jtreg/compiler/loopopts/superword/TestSplitPacks.java
>>    
>>    fix bug number
>>  - cleanup
>>  - ... and 17 more: https://git.openjdk.org/jdk/compare/490825fb...c39ace04
>
> src/hotspot/share/opto/superword.cpp line 969:
> 
>> 967:   }
>> 968: 
>> 969:   if (isomorphic(s1, s2) && !is_populate_index(s1, s2)) {
> 
> Note: As explained in `test7a`:
> If we pack the `iv + 1`, `iv + 2`, `iv + 3` ... `AddI` nodes, then we get a pack that is non-power-of-2 in most cases, since `iv + 0` has collapsed to `iv`, and is not an `AddI` but a `Phi` node. That then causes `split_packs_at_use_def_boundaries` to cut the use-packs, and things unravel from there.
> 
> So I had 2 options:
> 1. Create `Phi, AddI, AddI, AddI, ...` packs, and deal with the consequences of having nodes of different types in a pack. Plus, we would have the `iv` in a pack, which also has consequences all over the place... Not great.
> 2. Forbid anything that looks like a `PopulateIndex` to be packed. It does not need to be packed, and will still be vectorized, see `SuperWord::vector_opd`. I went with this option.

Agree with option 2.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/17848#discussion_r1503091498