RFR: 8236641: Improve Set.of(...).iterator() warmup characteristics

[Claes Redestad]

On 2020-01-18 17:51, Stuart Marks wrote:
> 
> 
> On 1/16/20 8:21 AM, Claes Redestad wrote:
>> I've cleaned up a bit (SALT can be made into a long to reduce casts
>> etc), and rolled it into the current patch:
>>
>> http://cr.openjdk.java.net/~redestad/8236641/open.01/
> 
> OK, I think this makes sense. But I did have to stare at the code for a 
> while to figure out what was going on. So, maybe a couple judicious 
> comments would be helpful.
> 
>    68         SALT32L = (int)((color * seed) >> 16) & 0xFFFF_FFFFL; // 
> avoid LSB and MSB
> 
> The comment "avoid LSB and MSB" isn't helping me here, and in fact it's 
> confusing because I thought it applied to the "& 0xFFFF_FFFFL" part. Is 
> there something else that ought to be said here, perhaps "select the 
> middle 32 bits of the product"?
> 
>   770                 idx = (int) ((SALT32L * elements.length) >>> 32);
> 1063                 idx = (int) ((SALT32L * (table.length >> 1)) >>> 
> 32) << 1;
> 
> These expressions replace floorMod(), which (at least to my brain) 
> idiomatically ensures the return value is in the range [0..len-1]. It's 
> not at all obvious what these do, although I think they do the same 
> thing. Can we add a comment something like, "add a 'random' starting 
> index based on the SALT"?

Thanks for reviewing!

Uploaded new version with improved comments here:

http://cr.openjdk.java.net/~redestad/8236641/open.01/

/Claes