RFR: 8316592: RISC-V: implement poly1305 intrinsic [v9]

[Hamlin Li]

On Wed, 8 Nov 2023 19:08:29 GMT, ArsenyBochkarev <duke at openjdk.org> wrote:

>> Hi everyone, please review this port of [AArch64](https://github.com/openjdk/jdk/blob/master/src/hotspot/cpu/aarch64/stubGenerator_aarch64.cpp#L7124) `_poly1305_processBlocks` intrinsic to RISC-V platform. 
>> 
>> ### Correctness checks
>> 
>> Tier 1 tests are passed. Also I explicitly ran the `test/jdk/com/sun/crypto/provider/Cipher/ChaCha20/unittest/Poly1305UnitTestDriver.java` test for multiple times and it passes.
>> 
>> ### Performance results on T-Head board
>> 
>> #### Results for enabled intrinsic:
>> 
>> Benchmark | (dataSize) | Mode | Cnt | Score | Error | Units
>> -- | -- | -- | -- | -- | -- | --
>> Poly1305DigestBench.digestBuffer | 64| thrpt | 3 | 247207.525 | 2853.920 | ops/s
>> Poly1305DigestBench.digestBuffer | 256 | thrpt | 3 | 221994.065 | 6891.601 | ops/s
>> Poly1305DigestBench.digestBuffer | 1024 | thrpt | 3 | 164485.375 | 4979.286 | ops/s
>> Poly1305DigestBench.digestBuffer | 16384 | thrpt | 3 | 27261.181 | 448.178 | ops/s
>> Poly1305DigestBench.digestBuffer | 1048576 | thrpt | 3 | 270.784 | 3445.077 | ops/s
>> Poly1305DigestBench.digestBytes | 64 | thrpt | 3 | 266049.018 | 9909.155 | ops/s
>> Poly1305DigestBench.digestBytes | 256 | thrpt | 3 | 231891.890 | 715.000 | ops/s
>> Poly1305DigestBench.digestBytes | 1024 | thrpt | 3 | 172746.932 | 1202.374 | ops/s
>> Poly1305DigestBench.digestBytes | 16384 | thrpt | 3 | 27626.478 | 341.915 | ops/s
>> Poly1305DigestBench.digestBytes | 1048576 | thrpt | 3 | 265.235 | 3522.458 | ops/s
>> Poly1305DigestBench.updateBytes | 64 | thrpt | 3 | 3394516.156 | 14656.687 | ops/s
>> Poly1305DigestBench.updateBytes | 256 | thrpt | 3 | 1463745.045 | 19608.937 | ops/s
>> Poly1305DigestBench.updateBytes | 1024 | thrpt | 3 | 459312.198 | 1720.655 | ops/s
>> Poly1305DigestBench.updateBytes | 16384 | thrpt | 3 | 30969.117 | 813.712 | ops/s
>> Poly1305DigestBench.updateBytes | 1048576 | thrpt | 3 | 300.773 | 3345.716 | ops/s
>> 
>> #### Results for disabled intrinsic:
>> 
>> Benchmark | (dataSize) | Mode | Cnt | Score | Error | Units
>> -- | -- | -- | -- | -- | -- | -- 
>> Poly1305DigestBench.digestBuffer | 64 | thrpt | 3 | 225424.813 | 1083.844 | ops/s
>> Poly1305DigestBench.digestBuffer | 256 | thrpt | 3 | 167848.372 | 3488.837 | ops/s
>> Poly1305DigestBench.digestBuffer | 1024 | thrpt | 3 | 81802.600 | 1839.218 | ops/s
>> Poly1305DigestBench.digestBuffer | 16384 | thrpt | 3 | 7781.049 | 1101.150 | ops/s
>> Poly1305DigestBench.digestBuffer | 1048576 | thrpt | 3 | 118.778 | 74.388 | ops/s
>> Poly1305DigestBench.digestBytes | 64 | thrpt | 3 | 23510...
>
> ArsenyBochkarev has updated the pull request incrementally with two additional commits since the last revision:
> 
>  - Rename poly1305-specific functions
>  - Add final conditional reduce

src/hotspot/cpu/riscv/stubGenerator_riscv.cpp line 4474:

> 4472: 
> 4473:     __ srli(tmp1, U_2, 2);
> 4474:     __ shadd(tmp1, tmp1, tmp1, tmp2, 2);

I'm afraid there might be an overflow issue here. 
Consider a smplifed case(8 bits per U): U_2==0b11111111, tmp1 will be 0b111111, tmp1*5 will overflow, and there is carry for it.
If I'm right, then the fix is to use 2 steps rather than one to reduce, please check the aarch64 version for details.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/16417#discussion_r1389558004