So I located the probable cause of the problems people reported with
kernels built with egcs: bogus asm statements in in_cksum.c.  At least
on my machine, this caused the two temporary registers (which were
erroneously in the input lists, rather than the output lists, of the
asms, and were initialized to the same value) to be overlapped on the
same hard register, causing obvious lossage.

I've fixed the code, and this problem should be gone.

(Why this managed to work for me for so long, I can only guess.)