overhaul optimized x86_64 memset asm

on most cpu models, "rep stosq" has high overhead that makes it
undesirable for small memset sizes. the new code extends the
minimal-branch fast path for short memsets from size 15 up to size
126, and shrink-wraps this code path. in addition, "rep stosq" is
sensitive to misalignment. the cost varies with size and with cpu
model, but it has been observed performing 1.5 times slower when the
destination address is not aligned mod 16. the new code thus ensures
alignment mod 16, but also preserves any existing additional
alignment, in case there are cpu models where it is beneficial.

this version is based in part on changes proposed by Denys Vlasenko.

x86_64/memset: avoid performing final store twice

The code does a potentially misaligned 8-byte store to fill the tail
of the buffer. Then it fills the initial part of the buffer
which is a multiple of 8 bytes.
Therefore, if size is divisible by 8, we were storing last word twice.

This patch decrements byte count before dividing it by 8,
making one less store in "size is divisible by 8" case,
and not changing anything in all other cases.
All at the cost of replacing one MOV insn with LEA insn.

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>

x86_64/memset: simple optimizations

"and $0xff,%esi" is a six-byte insn (81 e6 ff 00 00 00), can use
4-byte "movzbl %sil,%esi" (40 0f b6 f6) instead.

64-bit imul is slow, move it as far up as possible so that the result
(rax) has more time to be ready by the time we start using it
in mem stores.

There is no need to shuffle registers in preparation to "rep movs"
if we are not going to take that code path. Thus, patch moves
"jump if len < 16" instructions up, and changes alternate code path
to use rdx and rdi instead of rcx and r8.

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>

optimized memset asm for i386 and x86_64

the concept of both versions is the same; they differ only in details.
for long runs, they use "rep movsl" or "rep movsq", and for small
runs, they use a trick, writing from both ends towards the middle,
that reduces the number of branches needed. in addition, if memset is
called multiple times with the same length, all branches will be
predicted; there are no loops.

for larger runs, there are likely faster approaches than "rep", at
least on some cpu models. for 32-bit, it's unlikely that there is any
faster approach that does not require non-baseline instructions; doing
anything fancier would require inspecting cpu capabilities. for
64-bit, there may very well be faster versions that work on all
models; further optimization could be explored in the future.

with these changes, memset is anywhere between 50% faster and 6 times
faster, depending on the cpu model and the length and alignment of the
destination buffer.