Unfortunately shifts are slow when the shift count is renamed in this way.

That's because the value still has to make it to the shifter... and the renamer's adder output clearly takes a different and higher-latency path than the one which doesn't go through it, confirming my original suspicion that this is a limitation of the bypass/forwarding network.

Ultimately it seems they are just playing around with where things happen; the addition can be done here but the value ultimately needs to go there for the shift, or the addition and shift can both happen there. One could envision a future change that puts the shifter in the renamer too. They're really constrained by the laws of physics.