[Bug rtl-optimization/58195] New: Missed optimization opportunity when returning a conditional

[Bug rtl-optimization/58195] New: Missed optimization opportunity when returning a conditional

[warp at iki dot fi]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58195

            Bug ID: 58195
           Summary: Missed optimization opportunity when returning a
                    conditional
           Product: gcc
           Version: 4.7.2
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: warp at iki dot fi

When compiling the following function:

int a(int input)
{
    return input == 0 ? 0 : -input;
}

gcc is unable to see that the conditional and returning 0 can be optimized
away, producing:

        movl    %edi, %edx
        xorl    %eax, %eax
        negl    %edx
        testl   %edi, %edi
        cmovne  %edx, %eax
        ret

For the record, I found out the above when I was testing what gcc would do with
this function:

int a(int input)
{
    int value = 0;
    for(int n = input; n != 0; ++n)
        ++value;
    return value;
}

gcc is able to optimize that into the same asm code as above, but no further
(not even if the conditional is written explicitly, as written in the first
function above.)

[Bug rtl-optimization/58195] Missed optimization opportunity when returning a conditional

[bisqwit at iki dot fi]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58195

Joel Yliluoma <bisqwit at iki dot fi> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |bisqwit at iki dot fi

--- Comment #1 from Joel Yliluoma <bisqwit at iki dot fi> ---
Problem confirmed on "gcc (GCC) 4.9.0 20140303 (experimental)" (SVN version) in
both 32-bit and 64-bit mode using -Ofast.

For comparison, Clang++ produces this instead (even on -O1): 

        negl    %edi
        movl    %edi, %eax
        ret

GCC misses an optimization opportunity here.

[Bug tree-optimization/58195] Missed optimization opportunity when returning a conditional

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58195

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2014-03-11
          Component|rtl-optimization            |tree-optimization
     Ever confirmed|0                           |1
      Known to fail|                            |4.9.0

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
phiopt should notice

  <bb 2>:
  if (input_2(D) != 0)
    goto <bb 3>;
  else
    goto <bb 4>;

  <bb 3>:
  iftmp.0_3 = -input_2(D);

  <bb 4>:
  # iftmp.0_1 = PHI <iftmp.0_3(3), 0(2)>

and value-replace iftmp.0_1 by -input_2(D).

[Bug tree-optimization/58195] Missed optimization opportunity when returning a conditional

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58195

--- Comment #3 from Marc Glisse <glisse at gcc dot gnu.org> ---
This looks like the unary version of:
http://gcc.gnu.org/ml/gcc-patches/2014-03/msg00014.html

and something that could be auto-generated from a pattern like:
(cond (ne @0 0) (negate @0) 0) -> (negate @0)
;-)

[Bug tree-optimization/58195] Missed optimization opportunity when returning a conditional

[bisqwit at iki dot fi]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58195

--- Comment #4 from Joel Yliluoma <bisqwit at iki dot fi> ---
Still confirmed on GCC 10 (Debian 10-20200324-1) 10.0.1 20200324 (experimental)
[master revision
596c90d3559:023579257f5:906b3eb9df6c577d3f6e9c3ea5c9d7e4d1e90536]


Seems I lack the oomph to update the "confirmed" state of this report.

[Bug tree-optimization/58195] Missed optimization opportunity when returning a conditional

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58195

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Depends on|                            |25290
                 CC|                            |pinskia at gcc dot gnu.org
           Assignee|unassigned at gcc dot gnu.org      |pinskia at gcc dot gnu.org
             Status|NEW                         |ASSIGNED

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Mine,  After PR 25290, the below match.pd pattern will just work :).

(In reply to Marc Glisse from comment #3) 
> and something that could be auto-generated from a pattern like:
> (cond (ne @0 0) (negate @0) 0) -> (negate @0)


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=25290
[Bug 25290] PHI-OPT could be rewritten so that is uses match

[Bug tree-optimization/58195] Missed optimization opportunity when returning a conditional

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58195

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Note the loop based one is a little more complex due to the way GCC IR handles
overflow being wrapping or undefined; I am just going to fix the case where
overflow is defined as wrapping. The other case needs more work and maybe even
a huge change in the IR handling.

[Bug tree-optimization/58195] Missed optimization opportunity when returning a conditional

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58195

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #5)
> Mine,  After PR 25290, the below match.pd pattern will just work :).
> 
> (In reply to Marc Glisse from comment #3) 
> > and something that could be auto-generated from a pattern like:
> > (cond (ne @0 0) (negate @0) 0) -> (negate @0)

Actually that was implemented with r12-2041.

It is just the loop that is left.  We do get the optimization if we use -fwrapv
:).

I will handle the loop one for GCC 13.

[Bug tree-optimization/58195] Missed optimization opportunity when returning a conditional

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58195

--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Note the loop one looks like:
  if (input_4(D) != 0)
    goto <bb 3>; [89.00%]
  else
    goto <bb 4>; [11.00%]

  <bb 3> [local count: 105119324]:
  _1 = (unsigned int) input_4(D);
  _3 = -_1;
  value_2 = (int) _3;

  <bb 4> [local count: 118111600]:
  # value_10 = PHI <value_2(3), 0(2)>

---- CUT ----
Which avoids undefined overflow for INT_MIN.