Thanks for the reply. But I am still very puzzled about this. When mapping with R1 reads first and counting with HTSeq and stranded option = yes, I get zero counts for features on the + strand. With the stranded option set to reverse I get the counts. When mapping with R2 reads first we get the exactly opposite results. With cufflinks we get fpkm-values almost identical no matter the order of input files.
I have been trying to find more information about how stranded reads are handled by TopHat/cufflinks, but without luck.