Adding fourty-five thousand lines of new code to the LLVM tree is a patch that adds support for ARM's 64-bit architecture to LLVM. This target is currently experimental and not built by default, but requires a compile-time switch for enabling. However, there's already talk amongst developers about enabling 64-bit ARM support by default. This work was merged into mainline LLVM this morning and will thus be found in the LLVM 3.3 release due out in a couple months time.
This initial AArch64 LLVM compiler support covers Assembly for all scalar instructions except for the late addition CRC instructions, code generation features needed for C++0x and C99, compilation support for the small memory model, absolute and position-independent code support, GNU-style TLS, and debugging support.
The 64-bit ARM compiler support, however, has yet to undergo any performance tuning for the greatest compiler performance. With there still being a few months to go until LLVM 3.3 is released, this will likely be achieved before the AArch64 LLVM support is in any released version.
Another limitation is that there isn't any NEON support for the AArch64 experimental target because the developer says there was "an outbreak of batshit insanisty in [ARM's] legal department."
The initial AArch64/ARMv8 instruction set support for LLVM can be found by cloning the LLVM SVN/Git repository where this big patch landed today.
LLVM 3.3 should be a fantastic release since aside from the 64-bit ARM support will also be the AMD Radeon GPU back-end, likely the enabling by default of the new loop vectorizer, and many other features for this open-source compiler infrastructure.