Some people I know previously asked this question to one of the Swift developers. I believe the reasoning was along the lines of (very liberally paraphrased/interpreted): Even aside from the problem of bootstrapping the compiler's compilation to start, performant compilation on C++ is a very mature and well-explored problem space. C++ already compiles on a ton of systems, and maintaining a C++-to-Swift(bootstrap compiler)-to-Swift(final compiler) compilation process is added complexity and invites repeated work.
From a language design point of view, the things needed to get a language to the point where it could successfully and efficiently compile itself would skew or re-order priorities in a much different fashion than if they were to focus on building and iterating on an applications and systems language.
tl;dr: LLVM and Clang and their assorted toolsets do quite a bit of valuable work; reinventing that to serve the language they're building is a feedback loop is a downside that also removes some upside from the equation, too.
From a language design point of view, the things needed to get a language to the point where it could successfully and efficiently compile itself would skew or re-order priorities in a much different fashion than if they were to focus on building and iterating on an applications and systems language.
tl;dr: LLVM and Clang and their assorted toolsets do quite a bit of valuable work; reinventing that to serve the language they're building is a feedback loop is a downside that also removes some upside from the equation, too.