This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Implement cond and induction cond reduction w/o REDUC_MAX_EXPR


On Thu, 22 Jun 2017, Richard Biener wrote:

> On Wed, 21 Jun 2017, Richard Biener wrote:
> 
> > 
> > During my attempt to refactor reduction vectorization I ran across
> > the special casing of inital values for INTEGER_INDUC_COND_REDUCTION
> > and tried to see what it is about.  So I ended up implementing
> > cond reduction support for targets w/o REDUC_MAX_EXPR by simply
> > doing the reduction in scalar code -- while that results in an
> > expensive epilogue the vector loop should be reasonably fast.
> > 
> > I still didn't run into any exec FAILs in vect.exp with removing
> > the INTEGER_INDUC_COND_REDUCTION special case thus the following
> > patch.
> > 
> > Alan -- is there a testcase (maybe full bootstrap & regtest will
> > unconver one) that shows how this is necessary?
> 
> Testing has some fallout and I discovered what this special-case is
> about.  I'll commit the below testcase to exercise the case.

After fixing the typo (sorry) the testcase also shows a latent
wrong-code bug.  We use an induction vector starting at index 1
for COND_REDUCTION to allow for a special value of zero but it
seems the INTEGER_INDUC_COND_REDUCTION special-casing of the
default value relies on the value of zero being special but
in the testcase it is meaningful...

In fact, given that for INTEGER_INDUC_COND_REDUCTION the
values, while based on an induction variable, do not have to
match 1:1, we can modify the testcase to use last = 2*i; or
last = i - 1;  Which means we have to add a second vector
to specify whether a lane was set or not, basically use
the COND_REDUCTION code?

Richard.

> Richard.
> 
> 2017-06-22  Richard Biener  <rguenther@suse.de>
> 
> 	* gcc.dg/vect/pr65947-14.c: New testcase.
> 
> Index: gcc/testsuite/gcc.dg/vect/pr65947-14.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/pr65947-14.c	(nonexistent)
> +++ gcc/testsuite/gcc.dg/vect/pr65947-14.c	(working copy)
> @@ -0,0 +1,44 @@
> +/* { dg-require-effective-target vect_condition } */
> +
> +#include "tree-vect.h"
> +
> +extern void abort (void) __attribute__ ((noreturn));
> +
> +#define N 27
> +
> +/* Condition reduction with matches only in even lanes at runtime.  */
> +
> +int
> +condition_reduction (int *a, int min_v)
> +{
> +  int last = N + 96;
> +
> +  for (int i = 0; i < N; i++)
> +    if (a[i] > min_v)
> +      last = i;
> +
> +  return last;
> +}
> +
> +int
> +main (void)
> +{
> +  int a[N] = {
> +  47, 12, 13, 14, 15, 16, 17, 18, 19, 20,
> +  1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
> +  21, 22, 23, 24, 25, 26, 27
> +  };
> +
> +  check_vect ();
> +
> +  int ret = condition_reduction (a, 46);
> +
> +  /* loop should have found a value of 0, not default N + 96.  */
> +  if (ret != 0)
> +    abort ();
> +
> +  return 0;
> +}
> +
> +/* { dg-final { scan-tree-dump-times "LOOP VECTORIZED" 2 "vect" { xfail { ! vect_max_reduc } } } */
> +/* { dg-final { scan-tree-dump-times "condition expression based on integer induction." 4 "vect" } } */
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]