RFR: 8257531: Super word not applied to a loop of simple Buffer operations [v2]

[Roland Westrelin]

On Fri, 4 Dec 2020 07:42:21 GMT, Vladimir Kozlov <kvn at openjdk.org> wrote:

>> In Buffer case there is additional (loop invariant) load from java/nio/IntBuffer.offset field.
>> SuperWord did not handle loop's invariants in complex address expressions (only when vectorization is forced):
>> 
>> `AddP(addr, AddP(addr, addr, LShiftL(ConvI2L(CastII(AddI(AddI(iv_phi, invariant_LoadI), incr))), shift)), 16)`
>> 
>> Invariant reference is used to make sure all memory accesses in vectors use the same one. And only when alignment code is generated in pre-loop it is used for code generatin. But that code did not take into account that invariant have to be scaled if needed as in this example (invariant_LoadI << shift).
>> 
>> I propose to record scaling (left shift) for invariant when it is present and use it to compare invariants and in pre-loop alignment code generation. I also slightly modified tracing output code for invariants.
>> 
>> This allow vectorize Java code which uses Buffer.
>> 
>> I added new test based on @PaulSandoz example. It tests presence of new vectors and correctness of vectorized code.
>> I included case (ByteBuffer.allocateDirect()) which is not vectorized because it uses Unsafe access - SuperWord complains about CastX2P(AddL) nodes which it sees instead of AddP.  We may consider vectorizing such code too later.
>> 
>> Testing: tier1-4, precheckin-comp (-Xcomp), renaissance test
>
> Vladimir Kozlov has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Requires test run only on x86, x64 and aarch64

Looks good to me.

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

Marked as reviewed by roland (Reviewer).

PR: https://git.openjdk.java.net/jdk/pull/1618