[Bug rtl-optimization/21182] [8/9/10/11 Regression] gcc can use registers but uses stack instead

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=21182

--- Comment #31 from Richard Biener <rguenth at gcc dot gnu.org> ---
-fno-tree-ter improves things quite a bit.  With -DNAILED_REGS gimple doesn't
do much because we treat registers as memory here.

For trunk

-O2               has 52 spills
-O2 -fno-tree-ter has 35 spills
-O2 -fno-tree-ter -fschedule-insns has 74 spills
-O2 -fno-tree-ter -fschedule-insns -fsched-pressure has 18 spills
-O2 -fschedule-insns -fsched-pressure has 17 spills

to me this really hints at out-of-SSA producing a very bad initial
schedule, by TER but also likely due to folding & friends doing
random stmt placing (it's all a single BB).  I think we'd benefit
quite a bit with killing TER (doing all interesting bits pre-RTL
or via SSA RTL forwprop) and ordering SSA def expansion for optimal
register pressure.