[Bug rtl-optimization/105338] [12 Regression] Regression: jump or cmove generated for pattern (x ? CST : 0)

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105338

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
So, if I change passes.def
      NEXT_PASS (pass_sink_code, true /* unsplit edges */);
line
      NEXT_PASS (pass_sink_code, false /* unsplit edges */);
I get back the 11.x code.
Before ce1 pass, in that case say f looks like:
    8: flags:CCZ=cmp(r83:SI,0)
    9: pc={(flags:CCZ!=0)?L22:pc}
      REG_DEAD flags:CCZ
      REG_BR_PROB 375809644
   10: NOTE_INSN_BASIC_BLOCK 3
    5: r82:SI=0
      REG_DEAD r83:SI
      ; pc falls through to BB 5
   22: L22:
   21: NOTE_INSN_BASIC_BLOCK 4
    4: r82:SI=0x5
   15: L15:
   18: NOTE_INSN_BASIC_BLOCK 5
use (r82)
i.e. essentially
  if (tmp83 != 0)
    tmp82 = 5;
  else
    tmp82 = 0;
But with vanilla trunk, it is instead:
    7: flags:CCZ=cmp(r83:SI,0)
    8: pc={(flags:CCZ==0)?L10:pc}
      REG_DEAD flags:CCZ
      REG_BR_PROB 697932188
    9: NOTE_INSN_BASIC_BLOCK 4
    4: r83:SI=0x5
   10: L10:
   11: NOTE_INSN_BASIC_BLOCK 5
use (r83)
i.e.
  if (tmp83 != 0)
    tmp83 = 5;

Before fwprop1, the code looks roughly the same except for the swapped
branches,
so in C
  if (tmp83 != 0)
    tmp82 = 5;
  else
    tmp82 = tmp83;
vs.
  if (tmp83 == 0)
    tmp82 = tmp83;
  else
    tmp82 = 5;
I think fwprop1 only works on extended basic blocks and therefore turns the
latter into
  if (tmp83 == 0)
    tmp82 = 0;
  else
    tmp82 = 5;
and not the former.