Necromancing this thread for a FLAC-centric comparison on my 38 CDs:
What difference does it make to be able to to switch decorrelation per block, like FLAC can?

What I did, with kind help from a software-savvy developer:
Converted each CD file to four mono files:
left, right, mid (average!) and side (difference) - simply ignoring that side needs 17 bits, just assuming that there are so few samples wrong that they won't affect the compression ratio on a full day of music
Compressed these and the stereo CDDA file with flac.exe, all at same block size (4096) to keep the overhead penalty from two monos constant (around 15 MB).  Chose three basic settings: "-0b4096 to -2b4096" (fixed-predictors only), -5 and -8p.

fixed pred'rs	-5 variations	-8p variations
13 311 MB	12 599 MB	12 520 MB	1ch left files + 1ch right files, incurs a 15 MB extra overhead
12 790 MB	12 153 MB	12 055 MB	1ch mid files + 1ch side files. The savings 521/446/465 are largely due to one single near-mono album that makes 221 of the 521.
12 737 MB	12 058 MB	11 984 MB	2ch files "-1/-5M/-8pM", encoder keeps  L&R, L&S, M&S, R&S choice for some blocks on; w/o that extra "15 MB" overhead
12 705 MB	12 028 MB	11 955 MB	2ch files "-2/-5/-8p", where the encoder checks L&R, L&S, M&S, R&S for each block.

Going from "forced mid side" to -M saves the 15 MB-ish overhead difference, so the real savings are not the apparent 53 / 95 / 71, but rather 15 less.  
Savings from the -m over -M, appear like 32 / 30 / 29, and those are apples to apples.

There is a big "YMMV" over corpus of course. Surprised that the savings aren't bigger? In perspective: plain -7 and -8 are 11970 and 11963. The difference -5 to -8 is twice the difference -m over -M.