builtins.c:
Allow MD backend to prevent the optimization of a bcopy() or memmove() of
size 1 (the size being known at compile-time) into an inline mempcpy()
expansion, which will in turn expand into a byte load and store operation.
This expansion loses precious address alignment information at some point
(because everybody knows that you can read a byte from any address, right?),
and this loses bigtime on strict alignment platforms which lack the ability
to accesse bytes directly, such as alpha (unless compiling with -mbwx and
runnning on a BWX-capable cpu).
config/alpha:
Require alignment of local arrays on word boundaries, and enable
the builtins.c `one-byte memcpy' workaround.
ok espie@
ec58a1093d