I haven't benched the latest round of XSLTers, but the performance on XSLT has historically been pretty abysmal. Most of the time people (read "Cocoon") say they have solved the problem by caching the output, but not all output can be cached. I doubt that XSLT can be very fast when actually parsing XML (for the data and the style sheet), so presumably the biggest speed wins would come from caching the stylesheet as perl code, and from generating the data in a perl data structure that the XSLT processor understands instead of actual XML (thus skipping the parsing step). Are we there yet? And if we are, aren't we kind of re-inventing the wheel? Perl doesn't need XML to make generic data structures.

Your point about not being constrained to a linear mode is actually one of the things I hold against XSLT, because all of the HTML monkeys I know like to think about page design in a linear mode. They don't want to specify a style for the product price; they want to write some HTML and say "the price goes here." It's just more intutitive to non-programmer types.

I can see value in pipelining for working on data, but I would do all of that data mangling before I get to a page generation stage, so the template tool itself doesn't need to support it.

Anyway, I have happilly used XML for other things and I don't usually take stabs at XSLT, but since he asked for opinions... I try to keep my preferences from coloring my templating guide too much (which now desperately needs an update, with new versions of AxKit, TT, Apache::ASP, etc. out).