[Bug ipa/100034] New: missed optimization for dead code elimination at -O3 (vs. -O1, -Os, -O2)

[Bug ipa/100034] New: missed optimization for dead code elimination at -O3 (vs. -O1, -Os, -O2)

[zhendong.su at inf dot ethz.ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100034

            Bug ID: 100034
           Summary: missed optimization for dead code elimination at -O3
                    (vs. -O1, -Os, -O2)
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: ipa
          Assignee: unassigned at gcc dot gnu.org
          Reporter: zhendong.su at inf dot ethz.ch
                CC: marxin at gcc dot gnu.org
  Target Milestone: ---

[641] % gcctk -v
Using built-in specs.
COLLECT_GCC=gcctk
COLLECT_LTO_WRAPPER=/local/suz-local/software/local/gcc-trunk/libexec/gcc/x86_64-pc-linux-gnu/11.0.1/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with: ../gcc-trunk/configure --disable-bootstrap
--prefix=/local/suz-local/software/local/gcc-trunk --enable-languages=c,c++
--disable-werror --enable-multilib --with-system-zlib
Thread model: posix
Supported LTO compression algorithms: zlib
gcc version 11.0.1 20210411 (experimental) [master revision
1d54b138417:e83b9cf4549:936d500dfc17f58f2507ecd0f7f26e4f197052ee] (GCC) 
[642] % 
[642] % gcctk -O1 -S -o O1.s small.c
[643] % gcctk -O3 -S -o O3.s small.c
[644] % 
[644] % wc O1.s O3.s
  23   45  420 O1.s
  41   78  697 O3.s
  64  123 1117 total
[645] % 
[645] % grep foo O1.s
[646] % grep foo O3.s
        call    foo
[647] % 
[647] % cat small.c
extern void foo(void);
static int a, b, f, g;
static int d() {
  while (g)
    f = 0;
  while (1)
    foo();
}
static void c() { d (); }
void e() {
  while (b) {
    if (!a)
      continue;
    c();
  }
}
int main() {
  e();
  return 0;
}

[Bug ipa/100034] missed optimization for dead code elimination at -O3 (vs. -O1, -Os, -O2)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100034

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
            Version|unknown                     |11.0
           Keywords|                            |missed-optimization
                 CC|                            |hubicka at gcc dot gnu.org
   Last reconfirmed|                            |2021-04-12

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Looks related to PR99991 - we do an IPA SRA clone but fail to inline it and
thus we end up with

void d.isra ()
{
  int D.1980;
  int g.2_1;

  <bb 2> [local count: 10631108]:

  <bb 3> [local count: 96646437]:
  g.2_1 = 0;
  if (g.2_1 != 0)
    goto <bb 3>; [89.00%]
  else
    goto <bb 4>; [11.00%]

  <bb 4> [local count: 1073741824]:
  foo ();
  goto <bb 4>; [100.00%]

}

int main ()
{
  int a.0_2;
  int b.1_3;

  <bb 2> [local count: 59461674]:
  goto <bb 6>; [100.00%]

  <bb 3> [local count: 1014686025]:
  a.0_2 = a;
  if (a.0_2 == 0)
    goto <bb 4>; [99.96%]
  else
    goto <bb 5>; [0.04%]

  <bb 4> [local count: 1014280151]:
  // predicted unlikely by continue predictor.
  goto <bb 6>; [100.00%]

  <bb 5> [local count: 405874]:
  d.isra ();

  <bb 6> [local count: 1073741824]:
  b.1_3 = b;
  if (b.1_3 != 0)
    goto <bb 3>; [94.50%]
  else
    goto <bb 7>; [5.50%]

  <bb 7> [local count: 59055800]:
  return 0;

}

where we optimize main to 'return 0' but fail to elide the unused d.isra.

So also a dup of the cases where a late IPA function reclaim is missing.

[Bug ipa/100034] missed optimization for dead code elimination at -O3 (vs. -O1, -Os, -O2)

[luoxhu at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100034

--- Comment #2 from luoxhu at gcc dot gnu.org ---
(In reply to Richard Biener from comment #1)
> Looks related to PR99991 - we do an IPA SRA clone but fail to inline it and
> thus we end up with
> 
> void d.isra ()
> {
>   int D.1980;
>   int g.2_1;
> 
>   <bb 2> [local count: 10631108]:
> 
>   <bb 3> [local count: 96646437]:
>   g.2_1 = 0;
>   if (g.2_1 != 0)
>     goto <bb 3>; [89.00%]
>   else
>     goto <bb 4>; [11.00%]
> 
>   <bb 4> [local count: 1073741824]:
>   foo ();
>   goto <bb 4>; [100.00%]
> 
> }
> 
> int main ()
> {
>   int a.0_2;
>   int b.1_3;
> 
>   <bb 2> [local count: 59461674]:
>   goto <bb 6>; [100.00%]
> 
>   <bb 3> [local count: 1014686025]:
>   a.0_2 = a;
>   if (a.0_2 == 0)
>     goto <bb 4>; [99.96%]
>   else
>     goto <bb 5>; [0.04%]
> 
>   <bb 4> [local count: 1014280151]:
>   // predicted unlikely by continue predictor.
>   goto <bb 6>; [100.00%]
> 
>   <bb 5> [local count: 405874]:
>   d.isra ();
> 
>   <bb 6> [local count: 1073741824]:
>   b.1_3 = b;
>   if (b.1_3 != 0)
>     goto <bb 3>; [94.50%]
>   else
>     goto <bb 7>; [5.50%]
> 
>   <bb 7> [local count: 59055800]:
>   return 0;
> 
> }
> 
> where we optimize main to 'return 0' but fail to elide the unused d.isra.
> 
> So also a dup of the cases where a late IPA function reclaim is missing.

early_inliner inlines e to main in -O3 due to param_early_inlining_insns is 14
for O3, but it is 6 for -O2, so want_early_inline_function_p returns different.

Then ipa-inline fails to inline d.isra by inline_functions_called_once as it is
called by two callees e->d.isra and main->d.isra.

But The two d.isra calls are removed by gimple 102t.ccp2 pass after all ipa
passes:


pr100034.O3.c.103t.objsz1:

;; Function d.isra (d.isra.0, funcdef_no=4, decl_uid=2014, cgraph_uid=7,
symbol_order=10) (executed once)

void d.isra ()
{
  int D.2016;

  <bb 2> [local count: 10631108]:

  <bb 3> [local count: 1073741824]:
  foo ();
  goto <bb 3>; [100.00%]

}



;; Function e (e, funcdef_no=2, decl_uid=1994, cgraph_uid=3, symbol_order=6)

void e ()
{
  <bb 2> [local count: 59461674]:
  return;

}



;; Function main (main, funcdef_no=3, decl_uid=1999, cgraph_uid=4,
symbol_order=7) (executed once)

int main ()
{
  <bb 2> [local count: 59461674]:
  return 0;

} 

Currently all IPA passes are run before gimple optimizations, is it possible to
run some passes like pass_rebuild_cgraph_edges and pass_ipa_remove_symbols
after some gimple optimisations expose new opertunities?