updated double_fast complementary insertion

[Cyan4973]

in a way which is more favorable to compression ratio, though slightly slower.

Detailed benchmark :
core i7-9700k (disabled turboboost), Ubuntu x64 19.04, gcc v8.3.0 :

filename	dev ratio	this patch	improv	c.speed	d.speed
silesia.tar	3.173	3.188	+0.47%	-1.9%	+0.28%
calgary.tar	3.075	3.098	+0.75%	-1.5%	+0.60%
dickens	2.767	2.789	+0.79%	-2.2%	+1.13%
mozilla	2.767	2.768	+0.04%	-2.5%	+0.02%
xml	8.364	8.407	+0.51%	-2.1%	+0.01%
book2	2.976	3.008	+1.08 %	-1.5%	+0.00%
progc	2.785	2.804	+0.68%	-2.6%	+0.32%
enwik8	2.806	2.826	+0.71%	-1.5%	+0.80%

Note : I must re-run speed tests, as it seems there is enough difference between cores to deserve pinning all measurements to the same core, for proper comparison.

edit : completed speed measurements.

edit 2 : final heuristic changed, see later benchmarks.

[Cyan4973]

Just completed speed measurements.

The impact on speed is a bit larger than I expected, being closer to 2% than 1%.
This is partially compensated by a very small improvement of decompression speed, but nothing huge.
As a consequence, the speed / ratio trade-off is more debatable (not bad, but not excellent either).

Should it be deemed not good enough, an alternative modification could be to change insertion of ip-2 by insertion of ip-1. In which case, compression is generally improved, but by very little, and at least compression speed should remain neutral.

[Cyan4973]

Some results from ip-1 insertion (instead of ip-2) :

filename	dev ratio	this patch	improv	c.speed	d.speed
silesia.tar	3.173	3.176	+0.09%	-0.25%	-0.65%
calgary.tar	3.075	3.077	+0.07%	-0.12%	-0.14%
dickens	2.767	2.764	-0.11%	-0.00%	-0.40%
mozilla	2.767	2.765	-0.07%	-0.84%	-0.57%
xml	8.364	8.351	-0.16%	-0.77%	-0.49%
book2	2.976	2.979	+0.10%
progc	2.785	2.790	+0.18%
enwik8	2.806	2.807	+0.04%

Outcome :

• Gains are still present, but are much smaller
• Gains are not guaranteed. It's generally positive, but not always.
• Compression speed is slightly slower, though that's negligible.
• Decompression speed is slightly slower too

This variant seems only marginally better than current "default". It doesn't look like a clear "must do".

[Cyan4973]

Measuring again this patch, but with clang v8.0 (instead of gcc v8.3.0) :

filename	dev ratio	this patch	improv	c.speed	d.speed
silesia.tar	3.173	3.188	+0.47%	-1.0%	+0.30%
calgary.tar	3.075	3.098	+0.75%	-0.9%	+0.04%
dickens	2.767	2.789	+0.79%	-2.1%	+0.70%
mozilla	2.767	2.768	+0.04%	-1.4%	+0.01%
xml	8.364	8.407	+0.51%	-1.3%	+0.03%
book2	2.976	3.008	+1.08 %	-0.76%	+0.55%
progc	2.785	2.804	+0.68%	-1.5%	-0.02%
enwik8	2.806	2.826	+0.71%	-1.3%	+0.75%

As can be seen, patch's results are more favorable with clang.
The compression speed reduction is much lower, making the trade-off with ratio more valuable.

Based on clang results, I'm inclined to consider the new trade-off better.

gcc results are more balanced, not clearly good nor bad, just "about meh-okay".

So, as a global summary, when adding these results, it feels more on the positive side.

[terrelln]

Results on an Intel i9-9900K with turboboost enabled pinned to CPU#0 compiled with gcc-9.1.0:

filename	dev ratio	this patch	improv	c.speed	d.speed
silesia.tar	3.173	3.188	+0.47%	-2.5%	+0.4%
calgary.tar	3.075	3.098	+0.75%	-2.6%	+1.3%
dickens	2.767	2.789	+0.79%
mozilla	2.767	2.768	+0.04%
xml	8.364	8.407	+0.51%
book2	2.976	3.008	+1.08 %
progc	2.785	2.804	+0.68%
enwik8	2.806	2.826	+0.71%	-2.7%	+1.5%

[Cyan4973]

Testing a new variation of complementary insertion, using the same nb of insertions as current, but organized differently (long at ip-2, short at ip-1).

As a consequence, the speed impact is negligible, barely measurable.
The compression ratio benefit is also much smaller, though generally positive.
However, this variant seems to introduce tiny decompression speed regressions.
(all these impacts are very small, quite close to noise level)

core i7-9700k (disabled turboboost), Ubuntu x64 19.04, gcc v9.1.0 :

filename	dev ratio	this patch	improv	c.speed	d.speed
silesia.tar	3.173	3.179	+0.19%	-0.19%	-0.22%
calgary.tar	3.075	3.078	+0.10%	-0.35%	-0.19%
enwik8	2.806	2.809	+0.11%	-0.33%	-0.29%
dickens	2.767	2.769	+0.07%	-0.07%	-0.24%
mozilla	2.767	2.768	+0.04%	-0.50%	-0.19%
xml	8.364	8.358	-0.07%	-0.14%	-0.11%
book2	2.976	2.981	+0.17%	0.00%	-0.27%
progc	2.785	2.791	+0.22%	+1.3%	-0.10%

edit : seems like an easier sell. Doesn't change the needle much, but is rather positive. Speed regression is insensible, so no bad effect for already deployed systems using level 3.

[Cyan4973]

Patch updated,
using latest trade-off (long at ip-2, short at ip-1).

[terrelln]

I see no compression speed regression on my machine with gcc-9.1.0, and see a small speed boost with clang.

Strangely, I can reproduce the decompression speed loss on gcc, but on clang I get a decompression speed boost.

[Cyan4973]

Updated _extDict variant.