[Bug ipa/96252] [10/11 Regression] mis-optimization where identical functions have very different codegen since gcc 10

[rguenth at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96252

Richard Biener  changed:

   What|Removed |Added

   Target Milestone|10.3|10.4

--- Comment #7 from Richard Biener  ---
GCC 10.3 is being released, retargeting bugs to GCC 10.4.

[Bug ipa/96252] [10/11 Regression] mis-optimization where identical functions have very different codegen since gcc 10

[hubicka at ucw dot cz via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96252

--- Comment #6 from Jan Hubicka  ---
Thinking of it, perhaps also inliner could take a hint that it is
inlining a tail call and do not produce unnecesary copy of the
functio parameter passed by value.

More generally, mod/ref has good chance to determine that parameter in
its original location is not modified by the call and we could avoid the
copy even for non-tailcalls?

Still would be interesting to know why copy propagation gives up.

[Bug ipa/96252] [10/11 Regression] mis-optimization where identical functions have very different codegen since gcc 10

[hubicka at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96252

Jan Hubicka  changed:

   What|Removed |Added

 Ever confirmed|0   |1
 Status|UNCONFIRMED |NEW
 CC||hubicka at gcc dot gnu.org
   Last reconfirmed||2021-02-14

--- Comment #5 from Jan Hubicka  ---
This looks like missed memory copy propagation to me.

We inline the icfed function back but for some reason we end up with all those
extra moves, so it does not seem to be problem with missed tailcall

IPA function summary for bool cmp_y(cmp, cmp)/767 inlinable
  global time: 92.095312
  self size:   13
  global size: 14
  min size:   11
  self stack:  0
  global stack:0
size:11.00, time:90.095312
size:3.00, time:2.00,  executed if:(not inlined)
  calls:
bool cmp_x(cmp, cmp)/804 inlined
  freq:1.00
  Stack frame offset 0, callee self size 0
  __lexicographical_compare_impl.isra/803 inlined
freq:1.00
Stack frame offset 0, callee self size 0

Funny thing is that inliner seems to believe it is going to reduce code size:
Considering bool cmp_x(cmp, cmp)/766 with 10 size
 to be inlined into bool cmp_y(cmp, cmp)/767 in unknown:0
 Estimated badness is -inf, frequency 1.00.
Badness calculation for bool cmp_y(cmp, cmp)/767 -> bool cmp_x(cmp,
cmp)/766
  size growth -3, time 16.00 unspec 18.00  big_speedup
  -inf: Growth -3 <= 0
  Adjusted by hints -inf

The body is:
bool cmp_y (struct cmp l, struct cmp r)
{
  int * __first1;
  int * __first2;
  struct cmp l;
  struct cmp r;
  int _8;
  int _9;
  bool _17;

   [local count: 1073741824]:
  l = l;
  r = r;
  goto ; [100.00%]

   [local count: 9416790681]:
  if (_8 > _9)
goto ; [3.66%]
  else
goto ; [96.34%]

   [local count: 9072136140]:
  __first1_11 = __first1_21 + 4;
  __first2_13 = __first2_2 + 4;
  if (  [(void *) + 256B] != __first2_13)
goto ; [95.91%]
  else
goto ; [4.09%]

   [local count: 9774538809]:
  # __first1_21 = PHI <__first1_11(4), _M_elems(2)>
  # __first2_2 = PHI <__first2_13(4), _M_elems(2)>
  _8 = MEM[(int *)__first1_21];
  _9 = MEM[(int *)__first2_2];
  if (_8 < _9)
goto ; [3.66%]
  else
goto ; [96.34%]

   [local count: 1073741824]:
  # _17 = PHI <0(3), 1(5), 0(4)>

   [local count: 1073741824]:
  # _17 = PHI <0(3), 1(5), 0(4)>
  l ={v} {CLOBBER};
  r ={v} {CLOBBER};
  return _17;

}

Richi,
in any case, we may want to avoid creating wrappers for functions with very
large parameters?

[Bug ipa/96252] [10/11 Regression] mis-optimization where identical functions have very different codegen since gcc 10

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96252

Richard Biener  changed:

   What|Removed |Added

   Target Milestone|10.2|10.3

--- Comment #4 from Richard Biener  ---
GCC 10.2 is released, adjusting target milestone.

[Bug ipa/96252] [10/11 Regression] mis-optimization where identical functions have very different codegen since gcc 10

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96252

Richard Biener  changed:

   What|Removed |Added

Summary|mis-optimization where  |[10/11 Regression]
   |identical functions have|mis-optimization where
   |very different codegen  |identical functions have
   |since gcc 10|very different codegen
   ||since gcc 10
   Target Milestone|--- |10.2
  Known to work||9.3.0
   Priority|P3  |P2

--- Comment #3 from Richard Biener  ---
GCC 9 correctly applies tail calling.