A64: Implement division in Lithium for smis

A64: Implement division in Lithium for smis

Implemented by splitting division into constant and non-constant RHS
instructions, and updating each for smi inputs.

BUG=

[jochen (gone - plz use gerrit), 2014-02-17 14:34:39]

I defer to Ulan

[Sven Panne, 2014-02-18 07:11:30]

Please delay this CL until I have landed
https://codereview.chromium.org/166793002/, which includes a real bug fix.

[Sven Panne, 2014-02-18 09:16:28]

As just discussed with Ulan offline, the master plan for this is as follows:
I'll land https://codereview.chromium.org/166793002/ first, keeping it
relatively minimal, and merge the fix back to 3.23 and 3.24 after having some
canary coverage. After that, I'd like to split off the part of div-i handling
constant power-of-2 divisors into a separate Lithium instruction on *all*
platforms. This was my initial plan, anyway, but I'll do it under a better name
than LDivConstI, which is a misnomer: It can't handle all divisors. Perhaps one
can tweak register allocation than a tiny bit, too.

When this is done, one can add Smi handling.

[Sven Panne, 2014-02-18 10:44:44]

After having a closer look at the code and some discussions with Benedikt, we
decided that the master plan should probably be slightly different: After
landing the CL with the fix, we should split both div-i and mod-i into 2 parts,
one for a constant right argument and one for a non-constant one. The constant
version can consistently handle power-of-2 cases and replace division by a
multiplication with the reciprocal.

Hopefully we can merge MathFloorOfDiv into this, it is only slightly different.
The current code as really a total mess with tons of duplicated code and
probably various hidden bugs (which have already been fixed at other places).
Furthermore it misses quite a few optimization opportunities, not only on ARM,
but on all platforms. :-/

[m.m.capewell, 2014-02-19 14:36:20]

OK, this sounds good. I'll close this review.