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

[jakub at gcc dot gnu.org]

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

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Ok, so at least for the 2nd function, the problem is that we reject it from
noce_try_cmove as well from noce_try_cmove_arith based on costs.
That is the case for the NEXT_PASS (pass_sink_code, false /* unsplit edges */);
compilation too, though it seems the cost is 4 higher in the vanilla trunk
case.

I think that can be perhaps fixed with:
--- gcc/config/i386/i386-expand.cc.jj   2022-04-22 14:18:27.000000000 +0200
+++ gcc/config/i386/i386-expand.cc      2022-04-22 15:13:47.263829089 +0200
@@ -3224,8 +3224,7 @@ ix86_expand_int_movcc (rtx operands[])
                }
              diff = ct - cf;

-             if (reg_overlap_mentioned_p (out, op0)
-                 || reg_overlap_mentioned_p (out, op1))
+             if (reg_overlap_mentioned_p (out, compare_op))
                tmp = gen_reg_rtx (mode);

              if (mode == DImode)
- at least I don't really see the point of using yet another temporary when we
already emitted the comparison earlier and all we emit is compare_op and
assignment to out.

Anyway, with NEXT_PASS (pass_sink_code, false /* unsplit edges */); it succeeds
then with cond_move_process_if_block.  But it doesn't with vanilla trunk,
because it punts on:
4060          /* Don't try to handle this if the condition uses the
4061             destination register.  */
4062          if (reg_overlap_mentioned_p (dest, cond))
4063            return FALSE;
I'd say it is reasonable to punt on that, because the whole
cond_move_process_if_block is meant to handle multiple cmoves, not just one,
and we handle all of them with the same cond.  The only case that could be
handled is if it is the very last set in then_bb and else_bb is not present, or
if it is the last set in else_bb.  I must say I'm also quite surprised we don't
really check any costs in the cond_move_process_if_block and just blindly
assume it will always be a win, so it seems it happily handles even the cases
of a single dest assignment that earlier noce_* routines attempt (but those do
check the costs and in this case punt).