Performance regressions with -n and -l

[nh2]

Using git bisecton on a Celeron 220 running Ubuntu 10.04 x64, I found the following performance regressions:

c0ac813 - runs at 18 mb/s with -n, 11 with -l
9c00229 - this commit slows -n down from 18 to 14 mb/s
29b1666 - this commit slows -n down from 15 to 10 mb/s
017ec9e - this finally reduces -n to 7,5 mb/s, 6 with -l

This means cutting performance in halves with the 8 commits between c0ac813 and 017ec9e !

[ckolivas]

Thanks, interesting. I've reviewed all the commits you described and every single one of them that reduces speed seems to do so by making the memory windows ever smaller. Unfortunately all these changes have been put in there to prevent lrzip failing by running out of ram (which was very easy to do). What is the ram configuration of your machine (it's obviously 32 bits), what type of file were you compressing, and how big was the file you used for the testing?