Bug 114678 - FAIL: gcc.dg/tree-ssa/range-sincos.c scan-tree-dump-not evrp "link_error" on s390
Summary: FAIL: gcc.dg/tree-ssa/range-sincos.c scan-tree-dump-not evrp "link_error" on ...
Status: NEW
Alias: None
Product: gcc
Classification: Unclassified
Component: tree-optimization (show other bugs)
Version: 14.0
: P3 normal
Target Milestone: ---
Assignee: Not yet assigned to anyone
URL:
Keywords: missed-optimization, testsuite-fail
Depends on:
Blocks: VRP
  Show dependency treegraph
 
Reported: 2024-04-10 13:39 UTC by Stefan Schulze Frielinghaus
Modified: 2024-04-12 08:38 UTC (History)
2 users (show)

See Also:
Host:
Target: s390*-*-*
Build:
Known to work:
Known to fail:
Last reconfirmed: 2024-04-11 00:00:00


Attachments

Note You need to log in before you can comment on or make changes to this bug.
Description Stefan Schulze Frielinghaus 2024-04-10 13:39:28 UTC
FAIL: gcc.dg/tree-ssa/range-sincos.c scan-tree-dump-not evrp "link_error"

I assume this is due to the case that on s390 we have an isinf optab.  Thus, prior evrp we end up with

void bar (double x)
{
  int _1;
  double _5;

  <bb 2> :
  _5 = sin (x_4(D));
  if (_5 ord _5)
    goto <bb 3>; [INV]
  else
    goto <bb 5>; [INV]

  <bb 3> :
  _1 = __builtin_isinf (x_4(D));
  if (_1 != 0)
    goto <bb 4>; [INV]
  else
    goto <bb 5>; [INV]

  <bb 4> :
  link_error ();

  <bb 5> :
  return;

}

whereas, e.g., on x86-64 we have

void bar (double x)
{
  double _1;
  double _7;

  <bb 2> :
  _7 = sin (x_6(D));
  if (_7 ord _7)
    goto <bb 3>; [INV]
  else
    goto <bb 7>; [INV]

  <bb 3> :
  if (x_6(D) unord x_6(D))
    goto <bb 4>; [INV]
  else
    goto <bb 5>; [INV]

  <bb 4> :
  link_error ();

  <bb 5> :
  _1 = ABS_EXPR <x_6(D)>;
  if (_1 u<= 1.79769313486231570814527423731704356798070567525844996599e+308)
    goto <bb 7>; [INV]
  else
    goto <bb 6>; [INV]

  <bb 6> :
  link_error ();

  <bb 7> :
  return;

}

If this optimization is all about ABS_EXPR the following patch would work on s390x and x86-64:

@@ -24,20 +24,20 @@ bar (double x)
     {
       if (__builtin_isnan (x))
        link_error ();
-      if (__builtin_isinf (x))
+      if (__builtin_fabs (x) == __builtin_inf ())
        link_error ();
     }
 }

However, I'm not sure whether this is not also a missed optimization in case of an isinf optab?
Comment 1 Richard Biener 2024-04-10 16:48:52 UTC
I guess VRP should handle __builtin_isinf and friends.
Comment 2 Andrew Pinski 2024-04-10 16:51:19 UTC
(In reply to Richard Biener from comment #1)
> I guess VRP should handle __builtin_isinf and friends.

Like was posted here:
https://gcc.gnu.org/pipermail/gcc-patches/2024-March/648303.html ?
Comment 3 Stefan Schulze Frielinghaus 2024-04-11 13:30:10 UTC
Thanks for the pointer.  I can confirm that the patch fixes this PR and also fixes

FAIL: gcc.dg/tree-ssa/vrp-float-abs-1.c scan-tree-dump-not evrp "link_error"
Comment 4 Andrew Pinski 2024-04-11 21:12:49 UTC
Confirmed then. Maybe you should xfail it for s390 for GCC 14 at least and then unxfail when the patch is committed?
Comment 5 Stefan Schulze Frielinghaus 2024-04-12 08:38:02 UTC
Ok, done in https://gcc.gnu.org/pipermail/gcc-patches/2024-April/649367.html