On Mar 14, 2010, at 1:58 PM, Anton Korobeynikov wrote: >> why not introduce a new float16 LLVM type? Then you wouldn't need >> intrinsics. > I'll let Chris answer :) > The short answer is here: http://nondot.org/sabre/LLVMNotes/HalfFloat.txt Right. it's a storage format, not a computation format. -Chris