RFR: 8257483: C2: Split immediate vector rotate from RotateLeftV and RotateRightV nodes

[Vladimir Ivanov]

On Tue, 1 Dec 2020 07:05:19 GMT, Dong Bo <dongbo at openjdk.org> wrote:

> Currently, for all CPUs, if optimizing RotateLeftV and RotateRightV with match rules in AD files, we have to implement both immediate and variable versions.
> 
> On aarch64, with match rules for vector rotation, immediate vector rotatation can be optimized with shift+insert instructions (i.e. SLI/SRI, ~20% improvements with an initial implementation).
> However there woule be performance regression for variable version, due to SLI/SRI have no register version in NEON intruction set and there is no register version for right shift neither.
> The instructions for match rules of vector rotate variable should be:
>     mov w9, 32
>     dup v13.4s, w9
>     sub  v20.4s, v13.4S, v19.4s
>     # For a loop, default version only has code below,
>     # the code above (loop invariants) are put outside of a loop.
>     sshl v17.4s, v16.4s, v19.4s
>     neg v18.16b, v20.16b       # on aarch64, vector right shift is implemented as left shift by negative shift count
>     ushl v16.4s, v16.4s, v18.4s
>     orr v16.16b, v17.16b, v16.16b
> 
> With this patch, immediate vector rotation can be matched alone and optimized on CPUs like aarch64.
> 
> Verified with linux-x86_64-server-fastdebug tier1-3 and passed.
> Also added immediate vector rotation tests to micro `test/micro/org/openjdk/bench/java/lang/RotateBenchmark.java`.
> Tested the micro on a x86_64/aarch64 server and witnessed no regressions.

Why do you introduce new ideal nodes (`RotateLeftImmV`/`RotateRightImmV`)? You already extended `is_vector_rotate_supported` to take the shift argument into account.

Also, you mention aarch64, but I see only x86 changes.

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

PR: https://git.openjdk.java.net/jdk/pull/1532