Are (some) time dummies redundant if another variable controls for a part of the sample period? For an OLS regression, on the one hand, I have a dummy variable for each sample year (from 2000 to 2012). On the other hand, I have a binary variable that is 1 if observations refer to a concrete part of the sample period (2008-2009).
Should I keep or drop the dummy variables that refer to 2008 and 2009?
 A: You should (probably) keep the year-specific dummy variable.
Just so we have some concrete terminology, let's call the year-specific dummy variables $T$ and call the special 2008-2009 dummy variable $S$. Call your predictors or independent variables $X$ and call your response or dependent variable $Y$. Call your regression coefficients $\beta$.
The idea behind a block of variables like $T$ is that it accounts for things that are different year-to-year, which are not otherwise accounted for by your $X$ but that might could be associated with both $X$ and $Y$, in which case your $\beta$ estimate will be what the econometricians call "biased," in that they are ascribing influence to $X$ that is more properly due to $T$.
If it helps, you can think of $T$ as generating a different regression intercept for each year of your data, thereby "shifting" the regression line up or down each year. Call this model $M$
If you include $S$ as well, this constitutes an additional shift, but one that is the same across multiple years. Call this model $M'$.
So think about what happens if you include $S$ but exclude $T_{08}$ and $T_{09}$. Call this model $M''$. 
The key point is that this $M''$ is the same as adding a constraint to $M$, that the coefficient on $T_{08}$ be equal to the coefficient on $T_{09}$. That is, you are forcing the intercept in 2008 to be equal to the intercept in 2009, while the other intercepts are "free" as before.
The setup in $M''$ can be interpreted to mean "the baseline $Y$ in 2008 is the same as in 2009, but is not necessarily the same as in any other year." If that's what you want, then you should exclude $T_{08}$ and $T_{09}$.
To me, it sounds implausible and contrived. Consider instead the interpretation of $M'$: "the baseline $Y$ is different in every year, and something special happened in 2008-2009 that shifted the baseline further in those two years." This leaves open the possibility of year-specific effects, while still accounting for whatever special event happened in 2008-2009 that justifies the special indicator.
In light of your comments about overfitting: unless your data is very, very small I wouldn't worry about it. And if your data really is so small that you really can't afford to lose one more degree of freedom, then you have other problems.
A: You'll want to see if having the additional 08/09 variable adds substantially to your model. If it does not, you can drop it. By having the additional variable, you risk overfitting. Look into using AIC or BIC to judge the relative qualities of your models.  Because I do not know exactly what you're trying to do, the additional dummy variable may provide information not provided in the year variable, but I suspect that it won't.  However, it's always nice to demonstrate things quantitatively, and if anyone questions you, you can show them the information criterions. Both AIC and BIC are easily implemented in any statistical software, and if you would like to indicate which is your preference, I'm sure we can find some nice tutorials for you. A quick google turns up:
Brian O'Meara Lab
AIC and BIC in R
EDA in R
Hope that helps.
