[Bug tree-optimization/103723] New: Loop invariant motion pass failed to remove unused code from loop

[Bug tree-optimization/103723] New: Loop invariant motion pass failed to remove unused code from loop

[lingling.kong7 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103723

            Bug ID: 103723
           Summary: Loop invariant motion pass failed to remove unused
                    code from loop
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: lingling.kong7 at gmail dot com
  Target Milestone: ---

https://godbolt.org/z/b7x1s5Yrn

For this case, the instruction  `mov     ecx, eax` in bb `.L3` could remove
from loop body and could sink to '.L8' and '.L1'.


After ivopt pass the tree dump is:

uint32_t test (uint32_t buf_avail, 
const uint8_t * buf, const uint8_t * buf_back)
{
  sizetype ivtmp.7;
  uint32_t len_test;
  unsigned char _3;
  sizetype _4;
  unsigned char _6;
  sizetype _7;
  unsigned int _15;
  unsigned int _16;
  unsigned int _17;

  <bb 2> [local count: 114863530]:
  if (buf_avail_8(D) > 2)
    goto <bb 6>; [94.50%]
  else
    goto <bb 8>; [5.50%]

  <bb 8> [local count: 6317494]:
  goto <bb 5>; [100.00%]

  <bb 6> [local count: 108546036]: 
 _15 = buf_avail_8(D) + 4294967293;
  _7 = (sizetype) _15;
  _4 = _7 + 3;
  goto <bb 4>; [100.00%]

  <bb 3> [local count: 958878296]:
  _16 = (unsigned int) ivtmp.7_19;
  len_test_12 = _16 + 1;
  ivtmp.7_18 = ivtmp.7_19 + 1;
  if (ivtmp.7_18 != _4)
    goto <bb 7>; [94.50%]
  else
    goto <bb 9>; [5.50%]

  <bb 9> [local count: 52738306]:
  # len_test_21 = PHI <len_test_12(3)>
  goto <bb 5>; [100.00%]

  <bb 7> [local count: 906139990]:

  <bb 4> [local count: 1014686026]:
  # ivtmp.7_19 = PHI <ivtmp.7_18(7), 2(6)>
  _17 = (unsigned int) ivtmp.7_19;
  len_test_13 = _17;
  _3 = MEM[(const uint8_t *)buf_9(D) + ivtmp.7_19 * 1];
  _6 = MEM[(const uint8_t *)buf_back_11(D) + ivtmp.7_19 * 1];
  if (_3 == _6)
    goto <bb 3>; [94.50%]
  else
    goto <bb 10>; [5.50%]

  <bb 10> [local count: 55807731]:
  # len_test_20 = PHI <len_test_13(4)>

  <bb 5> [local count: 114863531]:
  # len_test_14 = PHI <len_test_21(9), len_test_20(10), 2(8)>
  return len_test_14;

}

len_test_13 is unused in loop body and could sink to bb 5.

[Bug tree-optimization/103723] zero extend not moved out of the loop after IV-OPTS

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103723

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2021-12-15
            Summary|Loop invariant motion pass  |zero extend not moved out
                   |failed to remove unused     |of the loop after IV-OPTS
                   |code from loop              |
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
           Severity|normal                      |enhancement

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
>For this case, the instruction  `mov     ecx, eax` in bb `.L3` could remove from loop body and could sink to '.L8' and '.L1'.

Right the problem is not what you referenced in the comment #0 though, it is
the following instead:
  _4 = (unsigned int) ivtmp.7_15;

That is a truncate from 64bit to 32bit.
There is no pass after IV-OPTS that sinks the zero extend out of the loop
though. LIM on the gimple level is not doing it for some reason ...

At Lim4 we have:
  <bb 4> [local count: 1014686026]:
  # ivtmp.7_15 = PHI <ivtmp.7_7(7), 2(6)>
  _4 = (unsigned int) ivtmp.7_15;
  len_test_13 = _4;
  _3 = MEM[(const uint8_t *)buf_9(D) + ivtmp.7_15 * 1];
  _6 = MEM[(const uint8_t *)buf_back_11(D) + ivtmp.7_15 * 1];
  if (_3 == _6)
    goto <bb 3>; [94.50%]
  else
    goto <bb 10>; [5.50%]

  <bb 10> [local count: 55807731]:
  # len_test_16 = PHI <len_test_13(4)>

I don't know enough about LIM if it was designed to handle non stores to sink
but maybe it should.

[Bug tree-optimization/103723] zero extend not moved out of the loop after IV-OPTS

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103723

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Actually this is more complex, I had missed the original add too.
the full loop:
 <bb 3> [local count: 958878296]:
  _18 = (unsigned int) ivtmp.7_15;
  len_test_12 = _18 + 1;
  ivtmp.7_7 = ivtmp.7_15 + 1;
  if (ivtmp.7_7 != _21)
    goto <bb 7>; [94.50%]
  else
    goto <bb 9>; [5.50%]

  <bb 9> [local count: 52738306]:
  # len_test_17 = PHI <len_test_12(3)>
  goto <bb 5>; [100.00%]

  <bb 7> [local count: 906139990]:

  <bb 4> [local count: 1014686026]:
  # ivtmp.7_15 = PHI <ivtmp.7_7(7), 2(6)>
  _4 = (unsigned int) ivtmp.7_15;
  len_test_13 = _4;
  _3 = MEM[(const uint8_t *)buf_9(D) + ivtmp.7_15 * 1];
  _6 = MEM[(const uint8_t *)buf_back_11(D) + ivtmp.7_15 * 1];
  if (_3 == _6)
    goto <bb 3>; [94.50%]
  else
    goto <bb 10>; [5.50%]

  <bb 10> [local count: 55807731]:
  # len_test_16 = PHI <len_test_13(4)>

We can't just sink the cast though as that would violate the closed-loop-SSA.
We would need to sink also:
  _18 = (unsigned int) ivtmp.7_15;
  len_test_12 = _18 + 1;

Which we do at the RTL level ....

[Bug tree-optimization/103723] zero extend not sunk out of the loop after IV-OPTS in the sink pass

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103723

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|zero extend not moved out   |zero extend not sunk out of
                   |of the loop after IV-OPTS   |the loop after IV-OPTS in
                   |                            |the sink pass

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Note there's pass_sink_code after IVOPTs which is what is designed to perform
sinking.  That is probably confused by seeing

  <bb 5> [local count: 1014686026]:
  # ivtmp.7_15 = PHI <ivtmp.7_7(7), 2(3)>
  _4 = (unsigned int) ivtmp.7_15;
  _3 = MEM[(const uint8_t *)buf_9(D) + ivtmp.7_15 * 1];
  _6 = MEM[(const uint8_t *)buf_back_11(D) + ivtmp.7_15 * 1];
  if (_3 == _6)
    goto <bb 4>; [94.50%]
  else
    goto <bb 10>; [5.50%]

  <bb 10> [local count: 55807731]:

  <bb 6> [local count: 114863531]:
  # len_test_14 = PHI <len_test_12(9), _4(10), 2(8)>
  return len_test_14;

where it doesn't consider inserting on edges because usually the pred works
fine but here it's a no-op sink and splitting the edge might be profitable
because its a loop exit.

So the sinking pass needs tweaking for such case (but _not_ split the edge
in case the PHI use is on a simple merge)