On Sat, Jul 08, 2017 at 11:15:48AM -0700, Ryan Tandy wrote:
The bug reproduces consistently when building with stretch's toolchain and running stretch's kernel, but I have not seen it at all with jessie's kernel *nor* in a jessie chroot on stretch's kernel. I have also reproduced it with clang as well as gcc.
The newer compilers generate POWER8 vector instructions e.g. mtvsrd/mfvsrd while the older ones do not. A kernel bug related to (or triggered by) these instructions would meet the conditions I described. Howard suggested we could be looking at a kernel bug around FP/VSR save/restore.
