Here's a draft of an improved i386 memset.s based on the principles
Denys Vlasenko and I discussed on his and my x86_64 versions. Compared
to the current code, it reduces entry/exit overhead, increases the
length supported in the non-rep-stosl path, and aligns the rep-stosl.

My tests don't measure the misalignment penalty, but even in the
aligned case the rep-stosl path is slightly faster (~5 cycles per run,
out of at least 64 cycles and the non-rep-stosl path is significantly
faster (e.g. 33 vs 51 cycles at size 16 and 40 vs 57 at size 32).

Empirically the byte-register-access/left-shift method of extending
the fill value to a word performs better than imul for me, but the
margin is very small (at most 1 cycle). Since we support much older
cpus (like actual 486) where imul could be really slow, I think this
is the right approach in principle too. I used imul in the rep-stosl
path but haven't tested whether it's faster there.

The non-rep-stosl path only goes up to size 62. I think sizes up to
126 could benefit from it, but the string of stores was getting really
long.

Correctness has not been tested so there may be stupid bugs.

Rich