On Sat, Oct 3, 2026 at 4:27 AM Andrea Pinski
<[email protected]> wrote:
>
> On Sat, Oct 3, 2026 at 1:13 AM Andrea Pinski
> <[email protected]> wrote:
> >
> > There are 2 patterns that are similar in match.pd for changing into
> > a smaller abs. One handles `abs((ext)a)` and one for `(trunc)abs((ext)a)`.
> > The one with the truncate afterwards does not use absu for integer types
> > which can introduce undefined behavior in some cases.
> > So let's restrict the second one to non integral types as the first one
> > will handle those always; the check for casting from unsigned is another
> > difference but the casting from unsigned is handled by a different abs
> > pattern for non-negative values.
> >
> > Bootstrapped and tested on x86_64-linux-gnu.
> >
> > PR tree-optimization/127689
> >
> > gcc/ChangeLog:
> >
> > * match.pd (`(trunc)abs((ext)a)`): Move right
> > after `abs((ext)a)` pattern. Restrict to
> > non integral types.
> >
> > gcc/testsuite/ChangeLog:
> >
> > * gcc.dg/torture/pr127689-1.c: New test.
> >
> > Signed-off-by: Andrea Pinski <[email protected]>
> > ---
> > gcc/match.pd | 22 +++++++++---------
> > gcc/testsuite/gcc.dg/torture/pr127689-1.c | 28 +++++++++++++++++++++++
> > 2 files changed, 39 insertions(+), 11 deletions(-)
> > create mode 100644 gcc/testsuite/gcc.dg/torture/pr127689-1.c
> >
> > diff --git a/gcc/match.pd b/gcc/match.pd
> > index 2d1161b860b..648bda9bf11 100644
> > --- a/gcc/match.pd
> > +++ b/gcc/match.pd
> > @@ -214,6 +214,17 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> > (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
> > (convert (absu:utype @0)))))
> >
> > +/* (trunc)abs (extend x) --> abs (x) for non-integral types. */
> > +(simplify
> > + (convert (abs (convert@1 @0)))
> > + (if (!ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
> > + && !HONOR_SNANS (@1)
> > + && types_match (type, TREE_TYPE (@0))
> > + && element_precision (type) < element_precision (TREE_TYPE (@1))
> > + && (!VECTOR_TYPE_P (type)
> > + || target_supports_op_p (type, ABS_EXPR, optab_default)))
> > + (abs @0)))
> > +
> > #if GIMPLE
> > /* Optimize (X + (X >> (prec - 1))) ^ (X >> (prec - 1)) into abs (X). */
> > (simplify
> > @@ -10340,17 +10351,6 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> > (IFN_COPYSIGN @0 { build_minus_one_cst (type); })
> > (IFN_COPYSIGN @0 { build_one_cst (type); })))))
> >
> > -/* (trunc)abs (extend x) --> abs (x) */
> > -(simplify
> > - (convert (abs (convert@1 @0)))
> > - (if (optimize
> > - && !HONOR_SNANS (@1)
> > - && types_match (type, TREE_TYPE (@0))
> > - && element_precision (type) < element_precision (TREE_TYPE (@1))
> > - && (!VECTOR_TYPE_P (type)
> > - || target_supports_op_p (type, ABS_EXPR, optab_default)))
> > - (abs @0)))
> > -
> > (for froms (BUILT_IN_FMAF BUILT_IN_FMA BUILT_IN_FMAL)
> > tos (IFN_FMA IFN_FMA IFN_FMA)
> > (simplify
> > diff --git a/gcc/testsuite/gcc.dg/torture/pr127689-1.c
> > b/gcc/testsuite/gcc.dg/torture/pr127689-1.c
> > new file mode 100644
> > index 00000000000..cc169744bf2
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.dg/torture/pr127689-1.c
> > @@ -0,0 +1,28 @@
> > +/* { dg-do run } */
> > +/* { dg-additional-options "-fno-tree-forwprop" } */
> > +/* PR tree-optimization/127689 */
> > +
> > +
> > +__attribute__((noipa))void check(int v)
> > +{
> > + if (v != -4)
> > + __builtin_exit(1);
> > +}
> > +int p;
> > +signed char s, v;
> > +signed char t = 0x80;
> > +__attribute__((noipa))
> > +void y(signed char w) {
> > + int t1;
>
> Pushed with a small update to the testcase (which I missed when I was
> moving from godbolt to the testsuite; and then forgot to test the
> testcase until right before pushing the commit).
Oh I forgot to say what the change was.
Initialize t1 to w.
That is:
int t1;
replaced with:
int t1 = w;
>
> Thanks,
> Andrea
>
> > + t1 = __builtin_abs(t1); //256, 0x80
> > + p = t1;
> > + s = p;
> > + signed char t2 = t1; // -256, 0x80
> > + t = t2 % 3; // -2
> > + v = 2 * t; // -4
> > + check(v);
> > +}
> > +int main()
> > +{
> > + y(0x80);
> > +}
> > --
> > 2.34.1
> >
