Even with the fix the code assumes that Lisp_Object is aligned at least 
as strictly as void *; this should be verified. Plus, if the alignments 
are the same Emacs can avoid a runtime check each time through the loop, 
which should help branch prediction. I installed the attached into 
master; if it fixes Ulrich's problem I think this should be backported 
to emacs-26. Thanks to Ulrich, Eli, and Andreas for reporting this and 
suggesting the fix.