[Bug tree-optimization/97000] New: [nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[Bug tree-optimization/97000] New: [nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97000

            Bug ID: 97000
           Summary: [nvptx, openacc] internal compiler error: in
                    nvptx_find_par, at config/nvptx/nvptx.c:3293
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: vries at gcc dot gnu.org
  Target Milestone: ---

When building an x86_64 + nvptx accelerator setup with the patch adding
libatomic for nvptx (
https://gcc.gnu.org/pipermail/gcc-patches/2020-September/553393.html ) , we run
into:
...
during RTL pass: mach
/home/vries/oacc/trunk/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/asyncwait-1.f90:
In function ‘MAIN__._omp_fn.0’:
/home/vries/oacc/trunk/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/asyncwait-1.f90:19:
internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293
0x160c390 nvptx_find_par
        /home/vries/oacc/trunk/source-gcc/gcc/config/nvptx/nvptx.c:3293
0x160c48a nvptx_find_par
        /home/vries/oacc/trunk/source-gcc/gcc/config/nvptx/nvptx.c:3320
0x160c48a nvptx_find_par
        /home/vries/oacc/trunk/source-gcc/gcc/config/nvptx/nvptx.c:3320
0x160c48a nvptx_find_par
        /home/vries/oacc/trunk/source-gcc/gcc/config/nvptx/nvptx.c:3320
0x160c48a nvptx_find_par
        /home/vries/oacc/trunk/source-gcc/gcc/config/nvptx/nvptx.c:3320
0x160c522 nvptx_discover_pars
        /home/vries/oacc/trunk/source-gcc/gcc/config/nvptx/nvptx.c:3344
0x1611046 nvptx_reorg
        /home/vries/oacc/trunk/source-gcc/gcc/config/nvptx/nvptx.c:5055
0x1059c88 execute
        /home/vries/oacc/trunk/source-gcc/gcc/reorg.c:4028
...

[Bug tree-optimization/97000] [nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97000

--- Comment #1 from Tom de Vries <vries at gcc dot gnu.org> ---
At asyncwait-1.xnvptx-none.mkoffload.179t.slsr, we have the valid:
...
  <bb 2> [local count: 87490071]:
  _14 = .UNIQUE (OACC_FORK, 0, 2);
  _75 = .GOACC_DIM_SIZE (0);
  _76 = .GOACC_DIM_POS (0);
  _77 = .GOACC_DIM_POS (2);
  _60 = _75 * 32;
  _61 = _60 + 63;
  _63 = _61 / _60;
  _44 = _63 * _76;
  _65 = _44 * 32;
  _19 = _65 + _77;
  _71 = _63 * 32;
  _72 = _19 + _71;
  _20 = MIN_EXPR <_72, 64>;
  if (_19 < _20)
    goto <bb 4>; [27.00%]
  else
    goto <bb 3>; [73.00%]


  <bb 3> [local count: 437450366]:
  .UNIQUE (OACC_JOIN, _14, 2);
  return;
...
and at asyncwait-1.xnvptx-none.mkoffload.181t.tracer, we have in addition:
...
  <bb 4> [local count: 373582614]:
  .UNIQUE (OACC_JOIN, _14, 2);
  return;
...
which causes the ICE.

There's some code in ignore_bb_p in tracer.c to stop this from happening.  But
it doesn't work, because it only checks for .UNIQUE in the last stmt in the bb,
while in bb3 the last is a return (and in bb2, the .UNIQUE is the first stmt).

[Bug tree-optimization/97000] [nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97000

--- Comment #2 from Tom de Vries <vries at gcc dot gnu.org> ---
Tentative patch:
...
diff --git a/gcc/tracer.c b/gcc/tracer.c
index 82ede722534..ec97eb51538 100644
--- a/gcc/tracer.c
+++ b/gcc/tracer.c
@@ -99,6 +99,12 @@ ignore_bb_p (const_basic_block bb)
         must be duplicated in its entirety or not at all.  */
       if (gimple_code (g) == GIMPLE_TRANSACTION)
        return true;
+    }
+
+  for (gimple_stmt_iterator gsi = gsi_start_bb (CONST_CAST_BB (bb));
+       !gsi_end_p (gsi); gsi_next (&gsi))
+    {
+      gimple *g = gsi_stmt (gsi);

       /* An IFN_UNIQUE call must be duplicated as part of its group,
         or not at all.  */
...

[Bug tree-optimization/97000] [nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97000

--- Comment #3 from Tom de Vries <vries at gcc dot gnu.org> ---
(In reply to Tom de Vries from comment #2)
> Tentative patch:

That fixes all the new libgomp FAILs.

[Bug tree-optimization/97000] [11 Regression][nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97000

Tobias Burnus <burnus at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2020-09-09
            Summary|[nvptx, openacc] internal   |[11 Regression][nvptx,
                   |compiler error: in          |openacc] internal compiler
                   |nvptx_find_par, at          |error: in nvptx_find_par,
                   |config/nvptx/nvptx.c:3293   |at
                   |                            |config/nvptx/nvptx.c:3293
             Status|UNCONFIRMED                 |NEW
           Keywords|                            |ice-on-valid-code
            Version|unknown                     |11.0
   Target Milestone|---                         |11.0
                 CC|                            |burnus at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org

--- Comment #4 from Tobias Burnus <burnus at gcc dot gnu.org> ---
Bisecting points to r11-3009-gfab77644842869adc8871e133e4c3f4c35b2b245
PR tree-optimization/96931 - clear ctrl-altering flag more aggressively

    2020-09-04  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/96931
            * tree-cfgcleanup.c (cleanup_call_ctrl_altering_flag): If
            there's a fallthru edge and no abnormal edge the call is
            no longer control-altering.
            (cleanup_control_flow_bb): Pass down the BB to
            cleanup_call_ctrl_altering_flag.

@Richard: See also tentative patch in comment 2

[Bug tree-optimization/97000] [11 Regression][nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97000

--- Comment #5 from Tom de Vries <vries at gcc dot gnu.org> ---
Right, I just found this:
...
      /* IFN_UNIQUE should be the last insn, to make checking for it            
         as cheap as possible.  */
      || (gimple_call_internal_p (stmt)
          && gimple_call_internal_unique_p (stmt)))
    gimple_call_set_ctrl_altering (stmt, true);
...

[Bug tree-optimization/97000] [11 Regression][nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97000

--- Comment #6 from Tom de Vries <vries at gcc dot gnu.org> ---
I wonder if this will work:
...
diff --git a/gcc/tree-cfgcleanup.c b/gcc/tree-cfgcleanup.c
index f8169eef781..79f716b9dbe 100644
--- a/gcc/tree-cfgcleanup.c
+++ b/gcc/tree-cfgcleanup.c
@@ -212,7 +212,9 @@ static void
 cleanup_call_ctrl_altering_flag (basic_block bb, gimple *bb_end)
 {
   if (!is_gimple_call (bb_end)
-      || !gimple_call_ctrl_altering_p (bb_end))
+      || !gimple_call_ctrl_altering_p (bb_end)
+      || (gimple_call_internal_p (bb_end)
+         && gimple_call_internal_unique_p (bb_end)))
     return;

   int flags = gimple_call_flags (bb_end);
...

[Bug tree-optimization/97000] [11 Regression][nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97000

--- Comment #7 from Tom de Vries <vries at gcc dot gnu.org> ---
(In reply to Tom de Vries from comment #6)
> I wonder if this will work:
> ...
> diff --git a/gcc/tree-cfgcleanup.c b/gcc/tree-cfgcleanup.c
> index f8169eef781..79f716b9dbe 100644
> --- a/gcc/tree-cfgcleanup.c
> +++ b/gcc/tree-cfgcleanup.c
> @@ -212,7 +212,9 @@ static void
>  cleanup_call_ctrl_altering_flag (basic_block bb, gimple *bb_end)
>  {
>    if (!is_gimple_call (bb_end)
> -      || !gimple_call_ctrl_altering_p (bb_end))
> +      || !gimple_call_ctrl_altering_p (bb_end)
> +      || (gimple_call_internal_p (bb_end)
> +         && gimple_call_internal_unique_p (bb_end)))
>      return;
>  
>    int flags = gimple_call_flags (bb_end);
> ...

Yep, that seems to do the trick.

[Bug tree-optimization/97000] [11 Regression][nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97000

--- Comment #8 from Tom de Vries <vries at gcc dot gnu.org> ---
This detects the problem earlier, in the host compiler:
...
diff --git a/gcc/omp-offload.c b/gcc/omp-offload.c
index 32c2485abd4..fce01af7682 100644
--- a/gcc/omp-offload.c
+++ b/gcc/omp-offload.c
@@ -1148,6 +1148,7 @@ oacc_loop_discover_walk (oacc_loop *loop, basic_block bb)
          break;

        case IFN_UNIQUE:
+         gcc_assert (stmt == last_stmt (CONST_CAST_BB (bb)));
          enum ifn_unique_kind kind
            = (enum ifn_unique_kind) (TREE_INT_CST_LOW
                                      (gimple_call_arg (call, 0)));
...

[Bug tree-optimization/97000] [11 Regression][nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97000

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
Meh, this way of forcing UNIQUE last to speedup lookup is a hack ... but yes,
your patch from comment#7 looks OK if you add

      /* IFN_UNIQUE should be the last insn, to make checking for it
         as cheap as possible.  */

before the new check to say what this is about.  Note this will re-introduce
the issue that passes cannot insert on loop entry edge without creating new BBs
even if we forced preheaders to be created.  For some very special case, when
UNIQUE is involved.  Thus we'll likely have even more hard to trigger latent
bugs (before it was also former setjmp calls).

[Bug tree-optimization/97000] [11 Regression][nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97000

--- Comment #10 from Tom de Vries <vries at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #9)
> Meh, this way of forcing UNIQUE last to speedup lookup is a hack ... but
> yes, your patch from comment#7 looks OK if you add
> 
>       /* IFN_UNIQUE should be the last insn, to make checking for it
>          as cheap as possible.  */
> 
> before the new check to say what this is about.  Note this will re-introduce
> the issue that passes cannot insert on loop entry edge without creating new
> BBs
> even if we forced preheaders to be created.  For some very special case, when
> UNIQUE is involved.  Thus we'll likely have even more hard to trigger latent
> bugs (before it was also former setjmp calls).

Filed PR97008 - "[openacc] Remove invariant that IFN_UNIQUE is last stmt in bb"

[Bug tree-optimization/97000] [11 Regression][nvptx, openacc] internal compiler error: in nvptx_find_par, at config/nvptx/nvptx.c:3293

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97000

Tom de Vries <vries at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |FIXED
           Keywords|                            |openacc

--- Comment #11 from Tom de Vries <vries at gcc dot gnu.org> ---
Patch committed (
https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=f96b6328fa7162111e4440c5901ef313ed3e6c9a
), no test-case required.

Marking resolved-fixed.