Huge pages

[paulbiss]

These are only turned on for internal kernels, we should enable them when supported externally.

[paulbiss]

I think I've got this partially working. I'm not seeing anything measurable from huge pages in the JIT, but I haven't gotten hugifyText working for hot runtime functions yet.

[wjywbs]

How's your progress? I replaced hash_string_cs_fallback with simply return hash_string_cs_unaligned_crc(arKey, nKeyLength); and saw 1-2% improvement. It would be beneficial to get copyHashFuncs working.

hhvm/hphp/util/hash.cpp

Line 48 in 4498e2b

strhash_t hash_string_i_fallback(const char *arKey, uint32_t nKeyLength) {

[paulbiss]

@wjywbs I had some issues with mapping the text section, I'll see if I can take another swing at it this week. Thanks for the update!

[markw65]

We should definitely get the huge page thing working, but we should also probably add a compilation option to just use the sse4 version of the crc code, rather than trying to dynamically patch it. Most people will either want it on, or want it off (rather than having a mix of machines, some with, and some without sse4).

[octmoraru]

@paulbiss What was the outcome of this effort? Were you able to make it work or was it dropped?

[paulbiss]

@octmoraru we bulk closed a bunch of lockdown issues, we haven't done this yet and we definitely should.

[lexidor]

I am going over old issues on this repository, to see which ones apply to the current versions of hhvm.

The last comment in #9063 confirms the huge page support:

(in order to make use of the optimization you need to configure the kernel to make huge pages available at boot time, and most users won't want this).