Which puts it into similar regions as Java, Go or D.
Unoptimized Lisp code can be a lot slower, but it then supports a bunch of interesting dynamic behavior.
The main approach of better performing Lisp systems is to provide a wide variety of programming styles and a lot of flexibility, while being at the same time fast enough, so that not much software needs to be written in C or assembler. For example in a large CAD system it is possible to write much of the code in Lisp - in a space where such an application has to compete with systems written in C and especially C++. Depending on the quality of the Lisp implementation, the C/C++ part gets larger or smaller.
For example one of PTC's Creo CAD systems is written mostly in Lisp - a bit more than 6 million lines of Lisp code was reported a few years back.
In most versions of Common Lisp you can see the exactly assembler output of your function, allowing you to optimise your code at any level you want. Like other people mentioned, you can add type declarations to ensure that the correct machine types are used, and you even have goto. What specific aspects of Lisp are you saying prevents optimisation? Especially given that Lisp has more low-level control than most other languages.
Obviously it isn't any closer to modern hardware than C, but as far from the hardware as the high-level parts are, the low-level parts are pretty reasonably close.