On Sat, 14 Sep 2019, Jakub Jelinek wrote:

> Hi!
> 
> As mentioned in the PR, the sqrt (x) < c optimization into x < c*c
> sometimes breaks the boundary case, if c2=c*c is inexact then in some cases
> we need to optimize it into x <= c*c rather than x < c*c.  The original
> bugreport is when c is small and c2 is 0.0, then obviously we need <= 0.0
> rather than < 0.0, but the testcase includes another example where it makes
> a difference, plus has a >= testcase too.
> 
> Bootstrapped/regtested on powerpc64le-linux, ok for trunk?

I was hoping Joseph might chime in here...  anyway, does this assume
round-to-nearest or does it work with round to +-Inf as well?  I
realize this all is under flag_unsafe_math_optimizations, but
this flag is notoriously underspecified...  So the question is
whether we should disable the transform if c*c isn't exact and
flag_rounding_math?  The transform also doesn't seem to guard
against isnan (c) (-funsafe-math-optimizations sets
-fno-trapping-math and -fno-signed-zeros but not -ffinite-math-only
or disables itself on -frounding-math)

Otherwise the patch looks OK to me.

Thanks,
Richard.


> 2019-09-13  Jakub Jelinek  <ja...@redhat.com>
> 
>       PR tree-optimization/91734
>       * generic-match-head.c: Include fold-const-call.h.
>       * match.pd (sqrt(x) < c, sqrt(x) >= c): Check the boundary value and
>       in case inexact computation of c*c affects comparison of the boundary,
>       turn LT_EXPR into LE_EXPR or GE_EXPR into GT_EXPR.
> 
>       * gcc.dg/pr91734.c: New test.
> 
> --- gcc/generic-match-head.c.jj       2019-07-20 21:02:09.296821929 +0200
> +++ gcc/generic-match-head.c  2019-09-12 10:52:33.091366624 +0200
> @@ -29,6 +29,7 @@ along with GCC; see the file COPYING3.
>  #include "cgraph.h"
>  #include "vec-perm-indices.h"
>  #include "fold-const.h"
> +#include "fold-const-call.h"
>  #include "stor-layout.h"
>  #include "tree-dfa.h"
>  #include "builtins.h"
> --- gcc/match.pd.jj   2019-09-11 21:50:54.933504293 +0200
> +++ gcc/match.pd      2019-09-12 11:12:11.150987786 +0200
> @@ -3541,56 +3541,71 @@ (define_operator_list COND_TERNARY
>         if x is negative or NaN.  Due to -funsafe-math-optimizations,
>         the results for other x follow from natural arithmetic.  */
>         (cmp @0 @1)))
> -     (if (cmp == GT_EXPR || cmp == GE_EXPR)
> +     (if (cmp == LT_EXPR || cmp == LE_EXPR || cmp == GT_EXPR || cmp == 
> GE_EXPR)
>        (with
>         {
> -         REAL_VALUE_TYPE c2;
> +      REAL_VALUE_TYPE c2;
> +      enum tree_code ncmp = cmp;
>        real_arithmetic (&c2, MULT_EXPR,
>                         &TREE_REAL_CST (@1), &TREE_REAL_CST (@1));
>        real_convert (&c2, TYPE_MODE (TREE_TYPE (@0)), &c2);
> +      /* See PR91734: if c2 is inexact and sqrt(c2) < c (or sqrt(c2) >= c),
> +         then change LT_EXPR into LE_EXPR or GE_EXPR into GT_EXPR.  */
> +      if ((cmp == LT_EXPR || cmp == GE_EXPR) && !REAL_VALUE_ISINF (c2))
> +        {
> +          tree c3 = fold_const_call (CFN_SQRT, TREE_TYPE (@0),
> +                                     build_real (TREE_TYPE (@0), c2));
> +          if (c3 == NULL_TREE || TREE_CODE (c3) != REAL_CST)
> +            ncmp = ERROR_MARK;
> +          else if (real_less (&TREE_REAL_CST (c3), &TREE_REAL_CST (@1)))
> +            ncmp = cmp == LT_EXPR ? LE_EXPR : GT_EXPR;
> +        }
>         }
> -       (if (REAL_VALUE_ISINF (c2))
> -     /* sqrt(x) > y is x == +Inf, when y is very large.  */
> -     (if (HONOR_INFINITIES (@0))
> -      (eq @0 { build_real (TREE_TYPE (@0), c2); })
> -      { constant_boolean_node (false, type); })
> -     /* sqrt(x) > c is the same as x > c*c.  */
> -     (cmp @0 { build_real (TREE_TYPE (@0), c2); }))))
> -     (if (cmp == LT_EXPR || cmp == LE_EXPR)
> -      (with
> -       {
> -              REAL_VALUE_TYPE c2;
> -      real_arithmetic (&c2, MULT_EXPR,
> -                       &TREE_REAL_CST (@1), &TREE_REAL_CST (@1));
> -      real_convert (&c2, TYPE_MODE (TREE_TYPE (@0)), &c2);
> -       }
> -       (if (REAL_VALUE_ISINF (c2))
> -        (switch
> -      /* sqrt(x) < y is always true, when y is a very large
> -         value and we don't care about NaNs or Infinities.  */
> -      (if (! HONOR_NANS (@0) && ! HONOR_INFINITIES (@0))
> -       { constant_boolean_node (true, type); })
> -      /* sqrt(x) < y is x != +Inf when y is very large and we
> -         don't care about NaNs.  */
> -      (if (! HONOR_NANS (@0))
> -       (ne @0 { build_real (TREE_TYPE (@0), c2); }))
> -      /* sqrt(x) < y is x >= 0 when y is very large and we
> -         don't care about Infinities.  */
> -      (if (! HONOR_INFINITIES (@0))
> -       (ge @0 { build_real (TREE_TYPE (@0), dconst0); }))
> -      /* sqrt(x) < y is x >= 0 && x != +Inf, when y is large.  */
> -      (if (GENERIC)
> -       (truth_andif
> -        (ge @0 { build_real (TREE_TYPE (@0), dconst0); })
> -        (ne @0 { build_real (TREE_TYPE (@0), c2); }))))
> -     /* sqrt(x) < c is the same as x < c*c, if we ignore NaNs.  */
> -     (if (! HONOR_NANS (@0))
> -      (cmp @0 { build_real (TREE_TYPE (@0), c2); })
> -      /* sqrt(x) < c is the same as x >= 0 && x < c*c.  */
> -      (if (GENERIC)
> -       (truth_andif
> -        (ge @0 { build_real (TREE_TYPE (@0), dconst0); })
> -        (cmp @0 { build_real (TREE_TYPE (@0), c2); })))))))))
> +       (if (cmp == GT_EXPR || cmp == GE_EXPR)
> +     (if (REAL_VALUE_ISINF (c2))
> +      /* sqrt(x) > y is x == +Inf, when y is very large.  */
> +      (if (HONOR_INFINITIES (@0))
> +       (eq @0 { build_real (TREE_TYPE (@0), c2); })
> +       { constant_boolean_node (false, type); })
> +      /* sqrt(x) > c is the same as x > c*c.  */
> +      (if (ncmp != ERROR_MARK)
> +       (if (ncmp == GE_EXPR)
> +        (ge @0 { build_real (TREE_TYPE (@0), c2); })
> +        (gt @0 { build_real (TREE_TYPE (@0), c2); }))))
> +     /* else if (cmp == LT_EXPR || cmp == LE_EXPR)  */
> +     (if (REAL_VALUE_ISINF (c2))
> +      (switch
> +       /* sqrt(x) < y is always true, when y is a very large
> +          value and we don't care about NaNs or Infinities.  */
> +       (if (! HONOR_NANS (@0) && ! HONOR_INFINITIES (@0))
> +        { constant_boolean_node (true, type); })
> +       /* sqrt(x) < y is x != +Inf when y is very large and we
> +          don't care about NaNs.  */
> +       (if (! HONOR_NANS (@0))
> +        (ne @0 { build_real (TREE_TYPE (@0), c2); }))
> +       /* sqrt(x) < y is x >= 0 when y is very large and we
> +          don't care about Infinities.  */
> +       (if (! HONOR_INFINITIES (@0))
> +        (ge @0 { build_real (TREE_TYPE (@0), dconst0); }))
> +       /* sqrt(x) < y is x >= 0 && x != +Inf, when y is large.  */
> +       (if (GENERIC)
> +        (truth_andif
> +         (ge @0 { build_real (TREE_TYPE (@0), dconst0); })
> +         (ne @0 { build_real (TREE_TYPE (@0), c2); }))))
> +      /* sqrt(x) < c is the same as x < c*c, if we ignore NaNs.  */
> +      (if (ncmp != ERROR_MARK && ! HONOR_NANS (@0))
> +       (if (ncmp == LT_EXPR)
> +        (lt @0 { build_real (TREE_TYPE (@0), c2); })
> +        (le @0 { build_real (TREE_TYPE (@0), c2); }))
> +       /* sqrt(x) < c is the same as x >= 0 && x < c*c.  */
> +       (if (ncmp != ERROR_MARK && GENERIC)
> +        (if (ncmp == LT_EXPR)
> +         (truth_andif
> +          (ge @0 { build_real (TREE_TYPE (@0), dconst0); })
> +          (lt @0 { build_real (TREE_TYPE (@0), c2); }))
> +         (truth_andif
> +          (ge @0 { build_real (TREE_TYPE (@0), dconst0); })
> +          (le @0 { build_real (TREE_TYPE (@0), c2); })))))))))))
>     /* Transform sqrt(x) cmp sqrt(y) -> x cmp y.  */
>     (simplify
>      (cmp (sq @0) (sq @1))
> --- gcc/testsuite/gcc.dg/pr91734.c.jj 2019-09-12 10:52:33.094366596 +0200
> +++ gcc/testsuite/gcc.dg/pr91734.c    2019-09-12 10:49:10.000000000 +0200
> @@ -0,0 +1,34 @@
> +/* PR tree-optimization/91734 */
> +/* { dg-do run } */
> +/* { dg-add-options ieee } */
> +/* { dg-additional-options "-ffast-math -O2 -std=gnu99" } */
> +
> +__attribute__((noipa)) int
> +foo (float x)
> +{
> +  return __builtin_sqrtf (x) < __FLT_MIN__;
> +}
> +
> +__attribute__((noipa)) int
> +bar (float x)
> +{
> +  return __builtin_sqrtf (x) < 0x1.2dd3d0p-65f;
> +}
> +
> +__attribute__((noipa)) int
> +baz (float x)
> +{
> +  return __builtin_sqrtf (x) >= 0x1.2dd3d0p-65f;
> +}
> +
> +int
> +main ()
> +{
> +  if (!foo (0.0f))
> +    __builtin_abort ();
> +  if (!bar (0x1.63dbc0p-130f))
> +    __builtin_abort ();
> +  if (baz (0x1.63dbc0p-130f))
> +    __builtin_abort ();
> +  return 0;
> +}
> 
>       Jakub
> 

-- 
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg,
Germany; GF: Felix Imendörffer; HRB 247165 (AG München)

Reply via email to