RFR: 8300257: C2: vectorization fails on some simple Memory Segment loops

[Vladimir Kozlov]

On Thu, 9 Mar 2023 10:48:06 GMT, Roland Westrelin <roland at openjdk.org> wrote:

> In the test case `testByteLong1` (that's extracted from a memory
> segment micro benchmark), the address of the store is initially:
> 
> 
> (AddP#204 base#195 base#195 (AddL#164 (ConvI2L#158 (CastII#157 (LshiftI#107 iv#101))) invar#163))
> 
> 
> (#numbers are node numbers to help the discussion).
> 
> `iv#101` is the `Phi` of a counted loop. `invar#163` is the
> `baseOffset` load.
> 
> To eliminate the range check, the loop is transformed into a loop nest
> and as a consequence the address above becomes:
> 
> 
> (AddP#204 base#195 base#195 (AddL#164 (ConvI2L#158 (CastII#157 (LShiftI#107 (AddI#326 invar#308 iv#321)))) invar#163))
> 
> 
> `invar#308` is some expression from a `Phi` of the outer loop.
> 
> That `AddP` is transformed multiple times to push the invariants out of loop:
> 
> 
> (AddP#568 base#195 (AddP#556 base#195 base#195 invar#163) (ConvI2L#158 (CastII#157 (AddI#566 (LShiftI#565 iv#321) invar#577))))
> 
> 
> then:
> 
> 
> (AddP#568 base#195 (AddP#847 (AddP#556 base#195 base#195 invar#163) (AddL#838 (ConvI2L#793 (LShiftL#760 iv#767)) (ConvI2L#818 (CastII#779 invar#577)))))
> 
> 
> and finally:
> 
> 
> (AddP#568 base#195 (AddP#949 base#195 (AddP#855 base#195 (AddP#556 base#195 base#195 invar#163) (ConvI2L#818 (CastII#809 invar#577))) (ConvI2L#938 (LShiftI#896 iv#908))))
> 
> 
> `AddP#855` is out of the inner loop. 
> 
> This doesn't vectorize because:
> 
> - there are 2 invariants in the address expression but superword only
>   support one (tracked by `_invar` in `SWPointer`)
> 
> - there are more levels of `AddP` (4) than superword supports (3)
> 
> To fix that, I propose to no longer track the address elements in
> `_invar`, `_negate_invar` and `_invar_scale` but instead to have a
> single `_invar` which is an expression built by superword as it
> follows chains of `addP` nodes. I kept the previous `_invar`,
> `_negate_invar` and `_invar_scale` as debugging and use them to check
> that what vectorized with the previous scheme still does.
> 
> I also propose lifting the restriction on 3 levels of `AddP` entirely.

My tier1-4,xcomp and stress testing passed.
I looked and changes and they seems fine.
May be we need to run performance testing too.

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

Marked as reviewed by kvn (Reviewer).

PR: https://git.openjdk.org/jdk/pull/12942