[Bug tree-optimization/66010] Missed optimization after inlining va_list parameter

["vries at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66010

--- Comment #2 from vries at gcc dot gnu.org ---
Created attachment 35460
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35460&action=edit
Demonstrator patch

I tracked down which passes are responsible for making the code for f1 and f2
look the same at the optimized dump.

pass_dominator rewrites the va_arg memory access to use the same argument as
the va_start:
...
  # .MEM_2 = VDEF <.MEM_1(D)>
  # USE = nonlocal escaped
  # CLB = nonlocal escaped { D.1844 } (escaped)
  __builtin_va_startD.1030 (&apD.1844, 0);
  # .MEM_3 = VDEF <.MEM_2>
  apD.1859 = &apD.1844;
  ap.7_10 = &apD.1844;
  # VUSE <.MEM_3>
  _11 = MEM[(struct  *)&apD.1844].gp_offsetD.2;
...

pass_sra manages to get rid of the vdef on 'ap_3 = &apD.1844':
...
  # .MEM_2 = VDEF <.MEM_1(D)>
  # USE = nonlocal escaped
  # CLB = nonlocal escaped { D.1844 } (escaped)
  __builtin_va_startD.1030 (&apD.1844, 0);
  ap_3 = &apD.1844;
  ap.7_10 = &apD.1844;
  # VUSE <.MEM_2>
  _11 = MEM[(struct  *)&apD.1844].gp_offsetD.2;
...

and pass_dce removes the dead code, leaving us with:
...
  # .MEM_2 = VDEF <.MEM_1(D)>
  # USE = nonlocal escaped
  # CLB = nonlocal escaped { D.1836 } (escaped)
  __builtin_va_startD.1030 (&apD.1836, 0);
  # VUSE <.MEM_2>
  _9 = apD.1836.gp_offsetD.2;
...

By squeezing these 3 passes in between the lowering and optimization parts in
pass_starg, we manage to get:
...
f2: va_list escapes 0, needs to save 8 GPR units and 0 FPR units.
...