[PATCH] Prefer small negative constants in InstCombine

[Tim Northover]

>> Just an update, it turned out to be a terrible idea. Linpack got 66% worse.
>
> Have you analyzed why?

Turned out to be random variation runs (well, it'd better be since the
two binaries were identical!).

Michael gave me some tips on improving consistency of the numbers,
which may or may not have helped but certainly made me check what the
test-suite says much more thoroughly. I've come to the conclusion that
the timing differences are far too small to show up, so I resorted to
code-size measurements.

The result was that only 3 binaries changed at all in the test-suite,
all code-size regressions. An alternative patch at the DAGCombiner
level had just 1 change, also a regression. I've not found a good
combined diff/disassembler to really investage what happened there.

Rather a depressing day, all round.

I'll see if I can come up with another approach that actually makes
things better.

Tim.