Re: [PATCH] move the (a-b) CMP 0 ? (a-b) : (b-a) optimization from fold_cond_expr_with_comparison to match

[Andrew Pinski <pinskia@gmail.com>]

On Mon, Jul 12, 2021 at 4:47 AM Richard Biener via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> On Sun, Jul 11, 2021 at 4:12 AM apinski--- via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > From: Andrew Pinski <apinski@marvell.com>
> >
> > This patch moves the (a-b) CMP 0 ? (a-b) : (b-a) optimization
> > from fold_cond_expr_with_comparison to match.
>
> So I searched and I guess these transforms are produced from
>
>   /* If we have A op 0 ? A : -A, consider applying the following
>      transformations:
>
>      A == 0? A : -A    same as -A
>      A != 0? A : -A    same as A
>      A >= 0? A : -A    same as abs (A)
>      A > 0?  A : -A    same as abs (A)
>      A <= 0? A : -A    same as -abs (A)
>      A < 0?  A : -A    same as -abs (A)
>
>      None of these transformations work for modes with signed
>      zeros.  If A is +/-0, the first two transformations will
>      change the sign of the result (from +0 to -0, or vice
>      versa).  The last four will fix the sign of the result,
>      even though the original expressions could be positive or
>      negative, depending on the sign of A.
>
>      Note that all these transformations are correct if A is
>      NaN, since the two alternatives (A and -A) are also NaNs.  */
>   if (!HONOR_SIGNED_ZEROS (type)
>       && (FLOAT_TYPE_P (TREE_TYPE (arg01))
>           ? real_zerop (arg01)
>           : integer_zerop (arg01))
>       && ((TREE_CODE (arg2) == NEGATE_EXPR
>            && operand_equal_p (TREE_OPERAND (arg2, 0), arg1, 0))
>              /* In the case that A is of the form X-Y, '-A' (arg2) may
>                 have already been folded to Y-X, check for that. */
>           || (TREE_CODE (arg1) == MINUS_EXPR
>               && TREE_CODE (arg2) == MINUS_EXPR
>               && operand_equal_p (TREE_OPERAND (arg1, 0),
>                                   TREE_OPERAND (arg2, 1), 0)
>               && operand_equal_p (TREE_OPERAND (arg1, 1),
>                                   TREE_OPERAND (arg2, 0), 0))))
> ...
>
> I wonder at which point we can remove the code from fold-const.c?

I have to double check if after an updated patch, if that code does
anything that match does not do.
I will do that before I submit an updated patch.

>
> Some comments inline below.
>
> > OK? Bootstrapped and tested on x86_64-linux-gnu.
> >
> > gcc/ChangeLog:
> >
> >         * match.pd ((A-B) CMP 0 ? (A-B) : (B - A)):
> >         New patterns.
> >
> > gcc/testsuite/ChangeLog:
> >
> >         * gcc.dg/tree-ssa/phi-opt-25.c: New test.
> > ---
> >  gcc/match.pd                               | 48 ++++++++++++++++++++--
> >  gcc/testsuite/gcc.dg/tree-ssa/phi-opt-25.c | 45 ++++++++++++++++++++
> >  2 files changed, 90 insertions(+), 3 deletions(-)
> >  create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/phi-opt-25.c
> >
> > diff --git a/gcc/match.pd b/gcc/match.pd
> > index 30680d488ab..aa88381fdcb 100644
> > --- a/gcc/match.pd
> > +++ b/gcc/match.pd
> > @@ -4040,9 +4040,7 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> >    (cnd (logical_inverted_value truth_valued_p@0) @1 @2)
> >    (cnd @0 @2 @1)))
> >
> > -/* abs/negative simplifications moved from fold_cond_expr_with_comparison,
> > -   Need to handle (A - B) case as fold_cond_expr_with_comparison does.
> > -   Need to handle UN* comparisons.
> > +/* abs/negative simplifications moved from fold_cond_expr_with_comparison.
> >
> >     None of these transformations work for modes with signed
> >     zeros.  If A is +/-0, the first two transformations will
> > @@ -4098,6 +4096,50 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> >         (convert (negate (absu:utype @0))))
> >         (negate (abs @0)))))
> >   )
> > +
> > + /* (A - B) == 0 ? (A - B) : (B - A)    same as (B - A) */
> > + (for cmp (eq uneq)
> > +  (simplify
> > +   (cnd (cmp (minus@0 @1 @2) zerop) @0 (minus@3 @2 @1))
> > +    (if (!HONOR_SIGNED_ZEROS (type))
> > +     @3))
> > +  (simplify
> > +   (cnd (cmp (minus@0 @1 @2) zerop) integer_zerop (minus@3 @2 @1))
>
> So that makes me think why integer_zerop?  'type' should then be
> integer and thus never HONOR_SIGNED_ZEROS.

yes that should be done.

>
> Don't we also need the inverted condition case for completeness?

Yes we should. Though for phiopt we don't.


>
> > +    (if (!HONOR_SIGNED_ZEROS (type))
> > +     @3))
> > +  (simplify
> > +   (cnd (cmp @1 @2) integer_zerop (minus@3 @2 @1))
>
> I think this needs to be (cmp:c @1 @2)

This is now actually handled already by r14-3606-g3d86e7f4a8ae so I removed it.
I will submit a new patch in the next few days for this too.

Thanks,
Andrew Pinski

>
> > +    (if (!HONOR_SIGNED_ZEROS (type))
> > +     @3))
> > + )
> > + /* (A - B) != 0 ? (A - B) : (B - A)    same as (A - B) */
> > + (for cmp (ne ltgt)
> > +  (simplify
> > +   (cnd (cmp (minus@0 @1 @2) zerop) @0 (minus @2 @1))
> > +    (if (!HONOR_SIGNED_ZEROS (type))
> > +     @0))
> > + )
> > + /* (A - B) >=/> 0 ? (A - B) : (B - A)    same as abs (A - B) */
> > + (for cmp (ge gt)
> > +  (simplify
> > +   (cnd (cmp (minus@0 @1 @2) zerop) @0 (minus @2 @1))
> > +    (if (!HONOR_SIGNED_ZEROS (type)
> > +        && !TYPE_UNSIGNED (type))
> > +     (abs @0))))
> > + /* (A - B) <=/< 0 ? (A - B) : (B - A)    same as -abs (A - B) */
> > + (for cmp (le lt)
> > +  (simplify
> > +   (cnd (cmp (minus@0 @1 @2) zerop) @0 (minus @2 @1))
> > +    (if (!HONOR_SIGNED_ZEROS (type)
> > +        && !TYPE_UNSIGNED (type))
> > +     (if (ANY_INTEGRAL_TYPE_P (type)
> > +         && !TYPE_OVERFLOW_WRAPS (type))
> > +      (with {
> > +       tree utype = unsigned_type_for (type);
> > +       }
> > +       (convert (negate (absu:utype @0))))
> > +       (negate (abs @0)))))
> > + )
> >  )
> >
> >  /* -(type)!A -> (type)A - 1.  */
> > diff --git a/gcc/testsuite/gcc.dg/tree-ssa/phi-opt-25.c b/gcc/testsuite/gcc.dg/tree-ssa/phi-opt-25.c
> > new file mode 100644
> > index 00000000000..0f0e3170f8d
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.dg/tree-ssa/phi-opt-25.c
> > @@ -0,0 +1,45 @@
> > +/* { dg-options "-O2 -fno-signed-zeros -fdump-tree-phiopt" } */
> > +int minus1(int a, int b)
> > +{
> > +  int c = a - b;
> > +  if (c == 0) c = b - a;
> > +  return c;
> > +}
> > +int minus2(int a, int b)
> > +{
> > +  int c = a - b;
> > +  if (c != 0) c = b - a;
> > +  return c;
> > +}
> > +int minus3(int a, int b)
> > +{
> > +  int c = a - b;
> > +  if (c == 0) c = 0;
> > +  else c = b - a;
> > +  return c;
> > +}
> > +int minus4(int a, int b)
> > +{
> > +  int c;
> > +  if (a == b) c = 0;
> > +  else
> > +    c = b - a;
> > +  return c;
> > +}
> > +int abs0(int a, int b)
> > +{
> > +  int c = a - b;
> > +  if (c <= 0) c = b - a;
> > +  return c;
> > +}
> > +int negabs(int a, int b)
> > +{
> > +  int c = a - b;
> > +  if (c >= 0) c = b - a;
> > +  return c;
> > +}
> > +
> > +/* The above should be optimized at phiopt1 except for negabs which has to wait
> > +  until phiopt2 as -abs is not acceptable in early phiopt.  */
> > +/* { dg-final { scan-tree-dump-times "if" 1  "phiopt1"  } } */
> > +/* { dg-final { scan-tree-dump-not "if" "phiopt2" } } */
> > --
> > 2.27.0
> >