If you want further details, you can visit Joshua Starmer youtube collection of videos

In this one, you can see details about how DESeq2 normalize RNA-Seq data from raw data

And in this one, how does it edgeR

In a nutshell, because you'll be able to apply statistical models to your values. When you normalize to FPKM you lose your ability to do so and end up with less robust statistical tests

The short answer is...the statisticians who wrote the software to do differential expression analysis have thought a lot about how best to normalize, and you should just let them do it. Those softwares all want raw counts as input so they can normalize, you should give them that, and let them do their thing. If you know what you are doing, and have compelling reasons to apply a different normalization function, the softwares have ways to allow you to do that, but a simple FPKM method is not going to be better than what the authors of the software have worked out.