Looping Through LLVM's New Loop Vectorizer

Premiering with LLVM 3.2, which will debut later this month, is an automatic loop vectorizer. I've already delivered benchmarks of LLVM's new automatic loop vectorizer, but here's more details on this new LLVM compiler feature.

The LLVM Loop Vectorizer is the second vectorizer for the compiler infrastructure after its Basic Block Vectorizer for dealing with straight-line code. The Loop Vectorizer is meant to widen instructions in an original loop for operating on multiple consecutive loop iterations. As mentioned in the earlier Phoronix articles on the subject, this feature isn't enabled by default with the forthcoming LLVM 3.2 but can be flipped on via the -mllvm -vectorize-loops compiler switches.

On the LLVM Blog is a new posting from yesterday that goes into detail about LLVM's Loop Vectorizer. The posting also mentions that for the LLVM 3.3 release in 2013 they will be aiming to have this vectorizer enabled by default. The blog post covers a few examples of how LLVM can now vectorize different complex loops.

The post also shares that there's more work going forward:

The Loop Vectorizer is a target independent IR-level optimization that depends on target-specific information from the different backends. It needs to select the optimal vector width and to decide if vectorization is worthwhile. Users can force a certain vector width using the command line flag "-mllvm -force-vector-width=X", where X is the number of vector elements. At the moment, only the X86 backend provides detailed cost information, while other targets use a less accurate method.

The work on the Loop Vectorizer is not complete and the vectorizer has a long way to go. We plan to add additional vectorization features such as automatic alignment of buffers, vectorization of function calls and support for user pragmas. We also plan to improve the quality of the generated code.

Read about this new LLVM 3.2 feature at blog.llvm.org. Also to be found in this month's LLVM/Clang/DragonEgg 3.2 release is better PowerPC compiler support, Polly improvements, and much more.