[Bug rtl-optimization/64164] [4.9/5 Regression] one more stack slot used due to one less inlining level

["law at redhat dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64164

--- Comment #22 from Jeffrey A. Law <law at redhat dot com> ---
Let's try to stay focused.  Killing copyrename seems like a fine thing to do as
long as the resulting debug info is good.  But that's independent of this BZ.

I still find myself wondering if leaving the "0" instead of "_10" in that PHI
node is a reasonable approach.  Certainly if I hack uncprop to leave things in
that state, I get the code we want.

And ISTM that uncprop ought to leave the constant alone if the SSA_NAME holding
the constant conflicts with any of the other SSA_NAMEs in the PHI node that may
potentially coalesce with the PHI result.  That captures pretty well the case
where the constant is better than an SSA_NAME.

In this particular case, we have:

  # _28 = PHI <0(2), _29(3), _29(7), _10(8), _29(6)>

When _28 and _10 coalesce, the result then conflicts with _29 during IRA/LRA
because of the extended lifetime of _10).  Thus the annoying copies created by
out-of-ssa can't be eliminated.

WIth the proposed change, we'd instead have:

  # _28 = PHI <0(2), _29(3), _29(7), 0(8), _29(6)>

While we won't coalesce _28/_29 during out-of-ssa, LRA will see the copies and
note that the associated pseudos don't conflict and ultimately assign them to
the same hard register and the annoying copies are gone.