[Bug c++/81410] [5/6/7/8 Regression] -O3 breaks code

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81410

--- Comment #8 from Richard Biener  ---
Author: rguenth
Date: Tue Jul 18 13:55:47 2017
New Revision: 250312

URL: https://gcc.gnu.org/viewcvs?rev=250312=gcc=rev
Log:
2017-06-18  Richard Biener  

PR tree-optimization/81410
* tree-vect-stmts.c (vectorizable_load): Properly adjust for
the gap in the ! slp_perm SLP case after each group.

* gcc.dg/vect/pr81410.c: New testcase.

Added:
trunk/gcc/testsuite/gcc.dg/vect/pr81410.c
Modified:
trunk/gcc/ChangeLog
trunk/gcc/testsuite/ChangeLog
trunk/gcc/tree-vect-stmts.c

[Bug c++/81410] [5/6/7/8 Regression] -O3 breaks code

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81410

--- Comment #7 from Richard Biener  ---
t.ii:25:19: note: === vect_analyze_data_ref_accesses ===
t.ii:25:19: note: Detected interleaving store _10->x and _10->y
t.ii:25:19: note: Detected interleaving load MEM[(const struct Foo &)_8].x and
MEM[(const struct Foo &)_8].y
t.ii:25:19: note: Detected interleaving store of size 2 starting with _10->x =
_37;
t.ii:25:19: note: Detected interleaving load of size 3 starting with _37 =
MEM[(const struct Foo &)_8].x;
t.ii:25:19: note: There is a gap of 1 elements after the group
...
t.ii:25:19: note: Final SLP tree for instance:
t.ii:25:19: note: node
t.ii:25:19: note:   stmt 0 _10->x = _37;
t.ii:25:19: note:   stmt 1 _10->y = _38;
t.ii:25:19: note: node
t.ii:25:19: note:   stmt 0 _37 = MEM[(const struct Foo &)_8].x;
t.ii:25:19: note:   stmt 1 _38 = MEM[(const struct Foo &)_8].y;

(note no load permutation)

t.ii:25:19: note: Loop contains SLP and non-SLP stmts
t.ii:25:19: note: Updating vectorization factor to 4
t.ii:25:19: note: vectorization_factor = 4, niters = 5

  _37 = MEM[(const struct Foo &)_8].x;
  vect__37.14_78 = MEM[(long int *)vectp.12_80];
  vectp.12_73 = vectp.12_80 + 16;
  vect__37.15_72 = MEM[(long int *)vectp.12_73];
  vectp.12_71 = vectp.12_73 + 16;
  vect__37.16_70 = MEM[(long int *)vectp.12_71];
  vectp.12_69 = vectp.12_71 + 16;
  vect__37.17_68 = MEM[(long int *)vectp.12_69];
  vectp.12_67 = vectp.12_69 + 32;
  _38 = MEM[(const struct Foo &)_8].y;

so the gap is accounted for in the wrong place once instead of twice as
required.

C testcase:

typedef __UINT64_TYPE__ uint64_t;
uint64_t x[24];
uint64_t y[16];
uint64_t z[8];

void __attribute__((noinline)) foo()
{
  for (int i = 0; i < 8; ++i)
{
  y[2*i] = x[3*i];
  y[2*i + 1] = x[3*i + 1];
  z[i] = 1;
}
}

int main()
{
  for (int i = 0; i < 24; ++i)
x[i] = i;
  foo ();
  for (int i = 0; i < 8; ++i)
if (y[2*i] != 3*i || y[2*i+1] != 3*i + 1)
  __builtin_abort ();
  return 0;
}

[Bug c++/81410] [5/6/7/8 Regression] -O3 breaks code

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81410

Richard Biener  changed:

   What|Removed |Added

   Keywords||wrong-code
   Priority|P3  |P2
 Status|NEW |ASSIGNED
   Assignee|unassigned at gcc dot gnu.org  |rguenth at gcc dot 
gnu.org
   Target Milestone|--- |5.5

--- Comment #6 from Richard Biener  ---
I'll have a looksee.

[Bug c++/81410] [5/6/7/8 Regression] -O3 breaks code

[glisse at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81410

--- Comment #5 from Marc Glisse  ---
Seems related to vectorization. These lines look suspicious:

  vect__37.14_78 = MEM[(long int *)_30];
  vect__37.15_72 = MEM[(long int *)_30 + 16B];
  vect__37.16_70 = MEM[(long int *)_30 + 32B];
  vect__37.17_68 = MEM[(long int *)_30 + 48B];
  MEM[(long int *)_28] = vect__37.14_78;
  MEM[(long int *)_28 + 16B] = vect__37.15_72;
  MEM[(long int *)_28 + 32B] = vect__37.16_70;
  MEM[(long int *)_28 + 48B] = vect__37.17_68;

where _30 is for b, _28 is for a, and I would expect to see gaps in the reads
from b (+24, +48, +72 instead of +16, +32 and +48). But I haven't checked, this
is only a first guess.

[Bug c++/81410] [5/6/7/8 Regression] -O3 breaks code

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81410

Martin Liška  changed:

   What|Removed |Added

 Status|UNCONFIRMED |NEW
   Last reconfirmed||2017-07-12
 CC||jason at gcc dot gnu.org,
   ||marxin at gcc dot gnu.org
Summary|O3 breaks code  |[5/6/7/8 Regression] -O3
   ||breaks code
 Ever confirmed|0   |1

--- Comment #4 from Martin Liška  ---
Confirmed, started with r209313.