Add prefetchCDictTables CCtxParam (+10-20% cold dict compression speed)

[embg]

Summary

• Adds a new CCtxParam (prefetchCDictTables)
• Exposes it in the largeNbDicts benchmark
• Adds it to fuzzers

Description of the optimization

In some situations, zstd uses CDict tables in-place rather than copying them into the working context. (See docs on ZSTD_dictAttachPref_e for details). In such situations, compression speed is seriously impacted when CDict tables are "cold" (outside CPU cache).

This PR adds a CCtxParam (prefetchCDictTables) which instructs zstd to prefetch CDict tables when they are used in-place (specifically in level 1-4 dictMatchState matchfinders). For sufficiently small inputs, the cost of the prefetch will outweigh the benefit. For sufficiently large inputs, zstd will by default memcpy() CDict tables into the working context, so there is no need to prefetch. This parameter is targeted at a middle range of input sizes, where a prefetch is cheap enough to be useful but memcpy() is too expensive.

The exact range of input sizes where this makes sense is best determined by careful experimentation (see below for measurements on one particular machine / dataset which demonstrate 10-20% wins for a particular working set size and input size). Rather than enabling this param for all inputs, the code which calls ZSTD_compress2() should use a size cutoff (tuned via experimentation) to select the best prefetch strategy for each input.

Measurements

I measured the effect of this param on the HTML dataset. I benchmarked on a Intel(R) Xeon(R) D-2191A CPU @ 1.60GHz machine with core isolation and turbo disabled.

We can see that the param is harmful for level3 even in the cold CDict scenario if the inputs are small enough (0-8K). For larger inputs (8-16K) at the same level, we see up to 20% wins. This demonstrates the need for selective application of this param.

[felixhandte]

LGTM. Just a couple nits.

[felixhandte]

Assuming tests pass, ship it!