[Bug tree-optimization/56608] [4.7/4.8 Regression] SLP seems to produce incorrect value with -ffast-math

2013-03-13 Thread rguenth at gcc dot gnu.org 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56608



Richard Biener rguenth at gcc dot gnu.org changed:



   What|Removed |Added



 Status|UNCONFIRMED |ASSIGNED

  Known to work||4.6.4

   Keywords||wrong-code

   Last reconfirmed||2013-03-13

 AssignedTo|unassigned at gcc dot   |rguenth at gcc dot gnu.org

   |gnu.org |

 Ever Confirmed|0   |1

Summary|SLP seems to produce|[4.7/4.8 Regression] SLP

   |incorrect value with|seems to produce incorrect

   |-ffast-math |value with -ffast-math

   Target Milestone|--- |4.7.3

  Known to fail||4.7.3, 4.8.0



--- Comment #3 from Richard Biener rguenth at gcc dot gnu.org 2013-03-13 
09:08:04 UTC ---

Confirmed.  Note we also generate non-optimal code in that we keep both

the vectorized and non-vectorized paths:



  _81 = (int) _80;

  uf_15(D)-autoCropHeight = _81;

...

  _84 = (double) _81;

  _85 = _83 / _84;

  if (aspectRatio_20  _85)

goto bb 9;

  else

goto bb 10;



  bb 9:

  _86 = _84 * aspectRatio_20;

  _87 = _86 + 5.0e-1;

  _88 = __builtin_ifloor (_87);



and vectorized:



  vect_p.26_224 = uf_15(D)-rotatedHeight;

  MEM[(struct ufraw_data *)vect_p.26_224] = vect_var_.22_217;



instead of re-loading the scalars from the vector.  That is, when

looking for scalar stores we start the SLP instance from we fail to

verify that no uses of the scalars in the instance remain after

vectorization.  Or if they remain, we fail to replace them with

vector extracts.



Not sure yet what's the wrong code issue (nothing obvious yet, still

investigating).

[Bug tree-optimization/56608] [4.7/4.8 Regression] SLP seems to produce incorrect value with -ffast-math

2013-03-13 Thread rguenth at gcc dot gnu.org 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56608



--- Comment #4 from Richard Biener rguenth at gcc dot gnu.org 2013-03-13 
10:43:23 UTC ---

Ick.  SLP replaces



  scale_51 = sqrt (_50);



with



  scale_51 = 0.0;



but has not eliminated all uses of it.  I have a simple fix, but really

I see no reason why DCE should not be able to remove the scalar calls given



  /* We only handle functions that do not read or clobber memory -- i.e.

 const or novops ones.  */

  if (!(gimple_call_flags (call)  (ECF_CONST | ECF_NOVOPS)))

return NULL_TREE;

[Bug tree-optimization/56608] [4.7/4.8 Regression] SLP seems to produce incorrect value with -ffast-math

2013-03-13 Thread rguenth at gcc dot gnu.org 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56608



--- Comment #5 from Richard Biener rguenth at gcc dot gnu.org 2013-03-13 
12:15:22 UTC ---

Author: rguenth

Date: Wed Mar 13 12:15:06 2013

New Revision: 196632



URL: http://gcc.gnu.org/viewcvs?root=gccview=revrev=196632

Log:

2013-03-13  Richard Biener  rguent...@suse.de



PR tree-optimization/56608

* tree-vect-slp.c (vect_schedule_slp): Do not remove scalar

calls when vectorizing basic-blocks.



* gcc.dg/vect/fast-math-bb-slp-call-3.c: New testcase.



Added:

trunk/gcc/testsuite/gcc.dg/vect/fast-math-bb-slp-call-3.c

Modified:

trunk/gcc/ChangeLog

trunk/gcc/testsuite/ChangeLog

trunk/gcc/tree-vect-slp.c