When we have a memory operand (HasSpillOperand() == true), and it
doesn't need a register immediately, split in an optimal position, which
is outside the outermost possible loop - just like Linear does.

This results in some modest improvements in perf, when compared
to baseline greedy. In particular Jetstream zlib x64: 4.66%, Life
(Emscripten x64) 11%; largest regression is in AreWeFastYet x64: 8%
and Corrections (Emsccripten x32) 10%

BUG=

Review URL: https://codereview.chromium.org/1306823005

Cr-Commit-Position: refs/heads/master@{#30498}