[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[law at redhat dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

Jeffrey A. Law  changed:

   What|Removed |Added

 Status|ASSIGNED|RESOLVED
 CC||law at redhat dot com
 Resolution|--- |FIXED

--- Comment #15 from Jeffrey A. Law  ---
Fixed by Richard's change on the trunk.

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

--- Comment #14 from Jeffrey A. Law  ---
Author: law
Date: Tue Feb 13 15:48:38 2018
New Revision: 257629

URL: https://gcc.gnu.org/viewcvs?rev=257629=gcc=rev
Log:
2018-02-12  Richard Sandiford  

gcc/
PR tree-optimization/84321
* tree-vrp.c (intersect_range_with_nonzero_bits): Fix VR_ANTI_RANGE
handling.  Also check whether the anti-range contains any values
that satisfy the mask; switch to a VR_RANGE if not.

gcc/testsuite/
PR tree-optimization/84321
* gcc.dg/pr84321.c: New test.

Added:
trunk/gcc/testsuite/gcc.dg/pr84321.c
Modified:
trunk/gcc/ChangeLog
trunk/gcc/testsuite/ChangeLog
trunk/gcc/tree-vrp.c

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

Richard Biener  changed:

   What|Removed |Added

   Priority|P3  |P1

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

--- Comment #13 from rsandifo at gcc dot gnu.org  
---
Testing a patch.  The VR_ANTI_RANGE code was wrong in a few ways. :-(

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

--- Comment #12 from rsandifo at gcc dot gnu.org  
---
(In reply to Jakub Jelinek from comment #11)
> (In reply to rsand...@gcc.gnu.org from comment #10)
> > I'd tried doing this in set_nonzero_bits first, before adding
> > intersect_range_with_nonzero_bits.  See
> > https://gcc.gnu.org/ml/gcc-patches/2018-02/msg00095.html for why it didn't
> > work.
> 
> IMHO that is a risk of all the VRP changes that some case will regress, the
> more important question is how much it will help in general, and I think it
> would.

Yeah I'd hoped so too, but...

> If it is only done in this special case, other passes can't benefit from the
> more accurate range info.

...when I did a comparison of the testsuite assembly output on a range
of targets, I didn't find any examples of things benefiting outside of
split_constant_offset, but quite a few regressions caused by trying
to interesect a VR_ANTI_RANGE with a reduced VR_RANGE.

The testsuite isn't the most representative codebase around, but still,
it suggests that we'd need to add some special cases elsewhere to
compensate.

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

--- Comment #11 from Jakub Jelinek  ---
(In reply to rsand...@gcc.gnu.org from comment #10)
> I'd tried doing this in set_nonzero_bits first, before adding
> intersect_range_with_nonzero_bits.  See
> https://gcc.gnu.org/ml/gcc-patches/2018-02/msg00095.html for why it didn't
> work.

IMHO that is a risk of all the VRP changes that some case will regress, the
more important question is how much it will help in general, and I think it
would.
If it is only done in this special case, other passes can't benefit from the
more accurate range info.
That said, such a change is likely inappropriate for stage4 at this point.

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

--- Comment #10 from rsandifo at gcc dot gnu.org  
---
(In reply to Jakub Jelinek from comment #8)
> This seems to fix it.  But as I said, the asserts still need to go.
> 
> --- gcc/tree-ssanames.c.jj2018-01-30 12:30:27.678359900 +0100
> +++ gcc/tree-ssanames.c   2018-02-12 11:32:51.768829381 +0100
> @@ -464,6 +464,21 @@ set_nonzero_bits (tree name, const wide_
>  }
>range_info_def *ri = SSA_NAME_RANGE_INFO (name);
>ri->set_nonzero_bits (mask);
> +  /* If it is a range, try to improve min/max from nonzero_bits.  */
> +  if (SSA_NAME_RANGE_TYPE (name) == VR_RANGE)
> +{
> +  wide_int msk = ri->get_nonzero_bits ();
> +  wide_int min = wi::round_up_for_mask (ri->get_min (), msk);
> +  wide_int max = wi::round_down_for_mask (ri->get_max (), msk);
> +  signop sgn = TYPE_SIGN (TREE_TYPE (name));
> +  if (wi::le_p (min, max, sgn)
> +   && wi::le_p (min, ri->get_max (), sgn)
> +   && wi::le_p (ri->get_min (), max, sgn))
> + {
> +   ri->set_min (min);
> +   ri->set_max (max);
> + }
> +}
>  }
>  
>  /* Return a widest_int with potentially non-zero bits in SSA_NAME

I'd tried doing this in set_nonzero_bits first, before adding
intersect_range_with_nonzero_bits.  See
https://gcc.gnu.org/ml/gcc-patches/2018-02/msg00095.html for why it didn't
work.

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

--- Comment #9 from Jakub Jelinek  ---
Perhaps we can handle anti range similarly and then not to
intersect_range_with_nonzero_bits at all.

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

--- Comment #7 from Jakub Jelinek  ---
int c;

void
foo (int *a, int b)
{
  int e;
  if (b == 1)
return;
  for (e = 0; e < (b & ~7); e += 8)
;
  for (++e; e < b;)
c = a[e];
}

actually.

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

Jakub Jelinek  changed:

   What|Removed |Added

 Status|SUSPENDED   |ASSIGNED

--- Comment #8 from Jakub Jelinek  ---
This seems to fix it.  But as I said, the asserts still need to go.

--- gcc/tree-ssanames.c.jj  2018-01-30 12:30:27.678359900 +0100
+++ gcc/tree-ssanames.c 2018-02-12 11:32:51.768829381 +0100
@@ -464,6 +464,21 @@ set_nonzero_bits (tree name, const wide_
 }
   range_info_def *ri = SSA_NAME_RANGE_INFO (name);
   ri->set_nonzero_bits (mask);
+  /* If it is a range, try to improve min/max from nonzero_bits.  */
+  if (SSA_NAME_RANGE_TYPE (name) == VR_RANGE)
+{
+  wide_int msk = ri->get_nonzero_bits ();
+  wide_int min = wi::round_up_for_mask (ri->get_min (), msk);
+  wide_int max = wi::round_down_for_mask (ri->get_max (), msk);
+  signop sgn = TYPE_SIGN (TREE_TYPE (name));
+  if (wi::le_p (min, max, sgn)
+ && wi::le_p (min, ri->get_max (), sgn)
+ && wi::le_p (ri->get_min (), max, sgn))
+   {
+ ri->set_min (min);
+ ri->set_max (max);
+   }
+}
 }

 /* Return a widest_int with potentially non-zero bits in SSA_NAME

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

rsandifo at gcc dot gnu.org  changed:

   What|Removed |Added

 Status|UNCONFIRMED |SUSPENDED
   Last reconfirmed||2018-02-12
   Assignee|unassigned at gcc dot gnu.org  |rsandifo at gcc dot 
gnu.org
 Ever confirmed|0   |1

--- Comment #6 from rsandifo at gcc dot gnu.org  
---
This should become a VR_RANGE rather than drop to VR_UNDEFINED.

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

--- Comment #5 from Jakub Jelinek  ---
Note, we have in set_range_info_raw:
  /* If it is a range, try to improve nonzero_bits from the min/max.  */
  if (range_type == VR_RANGE)
{
  wide_int xorv = ri->get_min () ^ ri->get_max ();
  if (xorv != 0)
xorv = wi::mask (precision - wi::clz (xorv), false, precision);
  ri->set_nonzero_bits (ri->get_nonzero_bits () & (ri->get_min () | xorv));
}
but don't have anything similar for VR_ANTI_RANGE (that is ok) and don't have
anything similar in set_nonzero_bits.

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

Jakub Jelinek  changed:

   What|Removed |Added

 CC||jakub at gcc dot gnu.org

--- Comment #4 from Jakub Jelinek  ---
Less reduced testcase so that it is valid:

int c;

void
foo (int *a, int b)
{
  int e;
  if (b == 1)
return -1;
  for (e = 0; e < (b & ~7); e += 8)
;
  for (++e; e < b;)
c = a[e];
}

This is because the range info and nonzero bits improvements improve just
gradually and improvements on nonzero bits don't have immediate effect on the
min/max.
We initially have:
  # RANGE [-2147483648, 2147483647] NONZERO 4294967288
  # e_11 = PHI 
  # RANGE [-2147483648, 2147483647] NONZERO 4294967289
  e_13 = e_11 + 1;
i.e. basically all range, but 0xfff8 and 0xfff9 masks.
Then vrp1 improves that to:
  # RANGE [0, 2147483647] NONZERO 2147483640
  # e_11 = PHI 
  # RANGE ~[-2147483647, 0] NONZERO 4294967289
  e_13 = e_11 + 1;
which is conservatively correct, but not perfect, because of the 0x7fff8
nonzero mask on the PHI guarantees that the actual range must be
[0, 2147483640].  And because e_11 is never INT_MAX, e_11 + 1 will never be
INT_MIN.  Later on ccp3 improves it further:
  # RANGE [0, 2147483647] NONZERO 2147483640
  # e_18 = PHI 
  # RANGE ~[-2147483647, 0] NONZERO 2147483641
  e_13 = e_18 + 1;
but just the nonzero mask.
So, while it would be nice to optimize this better, the asserts in
intersect_range_with_nonzero_bits are just wrong, just punt (return
VR_UNDEFINED) in those cases.

[Bug tree-optimization/84321] [8 Regression] ice in intersect_range_with_nonzero_bits, at tree-vrp.c:213

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84321

Andrew Pinski  changed:

   What|Removed |Added

   Keywords||ice-on-valid-code
  Component|c   |tree-optimization
   Target Milestone|--- |8.0
Summary|ice in  |[8 Regression] ice in
   |intersect_range_with_nonzer |intersect_range_with_nonzer
   |o_bits, at tree-vrp.c:213   |o_bits, at tree-vrp.c:213
   Severity|normal  |critical

--- Comment #3 from Andrew Pinski  ---
(In reply to David Binderman from comment #2)
> svn blame produced this for tree-vrp.c:213
> 
> 257491   rsandifo   gcc_checking_assert (wi::le_p (*min, *max, sgn));

r257491