As evidence goes, I don't have any either, but I Think the issue is related to the modularization of the kernel. all I really have are my own observations as I search for the perfect linux desktop for my personal use. What I do know is that on 3 machine the Boot time from power up to KDE desktop and then to open open office between Yoper i686 optimized, and CollegeLinux 2.5 with the same test (i486 optimized) is virtually identical. (feather linux was faster again, but its not really a full linux.) Mind you the hardware is not exactly cutting edge. PIII 700. a Duron 1100 and a AMD 2600+ (laptop) - There is also a p1 233, but you'll see why when I finish the shootout. Slack, fedora and mandrake and even SUSE were all slower, but due to more servers running and other basic Speed Tweaks rather than the settings the compiler used. That said, I think Yoper is an awesome desktop, CollegeLinux is also doing a lot of similar things, and other than perhaps mepis (debian based) these are what I tend to choose for a desktop distro. Its a real pity as I think Gnome 2.6 would be a brilliant desktop addition, but we will see. The shootout continues, and Yoper is getting put through its paces at the moment. But Servers are a different beast, and should be more concerned with stability over performance IMHO.
Anyway, I've seen it suggested that compiling the kernel and the c library for your particular machine and leaving all the other libraries/applications as generic i386 will make a small amount of difference, and optimising everything else won't make that much difference on top of that. Don't have any links or proof to back that up though :p
Just 'cos a distro is labelled as 'i386' doesn't mean it doesn't have optimisations for more recent processors in it, just that only use machine instructions that the i386 has.