This optimizes the passing of stack parameters in function calls.

For some architectures (ia32/x64), using pushes when possible instead
of bumping the stack and then storing parameters generates much
smaller code, and in some cases is faster (e.g. when a push of a memory
location can implement a memory-to-memory copy and thus elide an
intermediate load. On others (e.g. ARM), the benefit is smaller, where
it's only possible to elide direct stack pointer adjustment in certain cases
or combine multiple register stores into a single instruction in other limited
situations. On yet other platforms (ARM64, MIPS), there are no push instructions,
and this optimization isn't used at all.

Ideally, this mechanism would be used for both tail calls and normal calls,
but "normal" calls are currently pretty efficient, and tail calls are very
inefficient, so this CL sets the bar low for building a new mechanism to
handle parameter pushing that only needs to raise the bar on tail calls for now.

The key aspect of this change is that adjustment to the stack pointer
for tail calls (and perhaps later real calls) is an explicit step separate from
instruction selection and gap resolution, but aware of both, making it possible
to safely recognize gap moves that are actually pushes.

Review-Url: https://codereview.chromium.org/2082263002
Cr-Commit-Position: refs/heads/master@{#37477}