LLVM 3.3 was just released last month and there's still several months until the LLVM 3.4 release is expected, but already there's been changes like using the LLVM Loop Vectorizer at new optimization levels, performance changes, enabling the Radeon R600 LLVM back-end, and also talk of having the SLP Vectorizer on for -O3.
Anyhow, having out the Intel Core i7 990X system for some updated Linux tests, today I ran some LLVM Clang 3.2, 3.3, and 3.4 SVN 20130715 benchmarks for different C/C++ workloads. The results in full can be found on OpenBenchmarking.org within 1307158-SO-LLVM34COM27.
In one of the SciMark 2 tests there is already a great speed improvement out of LLVM/Clang 3.4.
Most workloads don't see any dramatic shifts yet on the in-development LLVM code, but be sure to see the rest of the results. You can also run your own compiler performance comparison using the Phoronix Test Suite.