Hi,

I have a customer asking for fast CRC64 support (specifically CRC64NVME) and 
would like to get some opinions on this.

Fast implementations leverage carryless multiplication instructions like 
`pmull` on aarch64, which are very difficult to match with C2. The Vector API 
isn't a good fit, as `pmull` doesn't perform vectorized operations across 
multiple lanes. As described in JDK-8379436, there isn't a standard way to 
return 128 bit values since the largest primitive is long at 64 bits, so adding 
carryless multiplication to Math requires a larger design.

Therefore I think there are two areas where the new algorithm could live.

1) java.util.zip.CRC64NVME

java.util.zip is the package where other checksums live, such as CRC32 and 
CRC32C, so based on existing precedent this is a good place to put it.

2) java.security.MessageDigest

MessageDigest is used for secure one way hash functions like MD5 and SHA. CRC64 
is a type of hashing function, so the existing API is a good fit although it 
isn't a secure one way function.

I understand it would need a CSR with a design and review, but do we think 
either of these is a path to pursue? Are there any other ideas on where this 
code could live?

Thanks,
Chad Rakoczy

Reply via email to