Comment # 2 on bug 82050 from Michel Dänzer

There was a recent change to LLVM which increased conformance with OpenCL
floating point semantics at some performance cost. That might explain at least
some of the difference.