https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84479
--- Comment #3 from kelvin at gcc dot gnu.org --- I had begun to look at this before Segher resolved it as invalid. One thing I had observed that their are some questionable practices in the rijndael-simd.cpp code. I had not yet dug deeply enough to confirm that these would actually cause problems. A patch that I merged with the trunk about the time this code began to fail changed the Power8 back-end to select more efficient instructions whenever vectors are "known" to be aligned. This patch also became more aggressive in its assessment of which vectors are aligned. In some cases, the instructions generated under the assumption that vectors are properly aligned in memory will "fail silently", producing undefined or unexpected results, when supplied with an address that is not properly aligned. While we were previously very conservative, we now consider that any pointer declared to refer to a vector is aligned on a multiple of the vector size. This means you cannot "safely" coerce "int *" to "vector int *", as happens, for example, around line 715 of rijndael-simd.cpp. There are various other places in this same source file that look suspicious to me.
