[eigen] Undesired formatting behaviour

[Johan Pauwels <johan.pauwels@xxxxxxxxxxxxx>]

Hi all,

I just found about Eigen this afternoon, but it looks exactly what I've 
been looking for to use in my PhD research. I've played around a bit 
with some basic examples of version 2.0.0 and I wanted to share what I 
think is a little inconsistency in the way the writing to an output 
stream works.

Not knowing the appropriate syntax yet, I first tried my luck with: cout 
<< setw(4) << matrix;
Since matrix is a Matrix object, I expected it to format all elements of 
the matrix with a minimum width of 4, but after a little digging in the 
code I found that the unexpected result was coming from the default 
empty matPrefix which was being set to width 4 and all the following 
elements not. Replacing int width = 0 by int width = s.width(); 
s.width(0); on line 129 of Eigen/src/Core/io.h gave me the result I 
intended. As far as I can see, this patch is compatible with the concept 
of IOFormat (of which I learned later on), since the default width of 0 
has no effect.

A similar, but more debatable case is that of ostream.precision(). If 
you specify the precision like cout << setprecision(3) << matrix; it 
gets ignored. The difference with the width case is that I cannot think 
of a straightforward patch that would add this functionality without 
changing the behaviour of IOFormat, since there is no "neutral" value 
for precision such as width(0) is. I guess the thing to do would be to 
initialise the default matPrefix to a to be defined special symbol that 
tells to take the current precision of the ostream when writing to that 
stream.

What I consider an inconsistency is that you use the IOFormat class at 
all to define the precision, since you still are dependent on the 
floatfield flag of the stream to use the precision you define in 
IOFormat. I mean, you still need to write cout << fixed << matrix; 
before the precision gets used with floats for example. Either you 
define all the formatting behaviour in the IOFormat class and inhibit 
the effects of the ostream flags or you read all ostream flags and act 
accordingly (where I'd prefer the latter). The AlignCols flag has added 
value compared to a simple width(x), so to put that into IOFormat makes 
sense to me in both cases. Either way, a trivial feature request is that 
I'd like setters for the variables in IOFormat such that you can 
initialise with the default IOFormat() format; and just change the one 
variable you want with format.setRowPrefix = "pre"; for example.

A last, completely unrelated thing is that I can find how to take the 
sum of a matrix, but not the product of the matrix. It could be that I'm 
looking over it, but if not, it's another feature request.

I hope this comes over as some constructive criticism as I intend it to 
be, because overall I think it is a terribly well done, feature-complete 
and elegant library.

Sincerely,
Johan Pauwels