On Thu, 2 May 2024, Manolis Tsamis wrote:
> On Thu, May 2, 2024 at 4:00 PM Richard Biener <rguent...@suse.de> wrote:
> >
> > On Tue, 23 Apr 2024, Manolis Tsamis wrote:
> >
> > > The original motivation for this pattern was that the following function
> > > does
> > > not fold to 'return 1':
> > >
> > > int foo(int *a, int j)
> > > {
> > > int k = j - 1;
> > > return a[j - 1] == a[k];
> > > }
> > >
> > > The expression ((unsigned long) (X +- C1) * C2) appears frequently as
> > > part of
> > > address calculations (e.g. arrays). These patterns help fold and simplify
> > > more
> > > expressions.
> > >
> > > PR tree-optimization/109393
> > >
> > > gcc/ChangeLog:
> > >
> > > * match.pd: Add new patterns for ((T)(A +- CST1)) * CST2 and
> > > ((T)(A +- CST1)) * CST2 + CST3.
> > >
> > > gcc/testsuite/ChangeLog:
> > >
> > > * gcc.dg/pr109393.c: New test.
> > >
> > > Signed-off-by: Manolis Tsamis <manolis.tsa...@vrull.eu>
> > > ---
> > >
> > > gcc/match.pd | 30 ++++++++++++++++++++++++++++++
> > > gcc/testsuite/gcc.dg/pr109393.c | 16 ++++++++++++++++
> > > 2 files changed, 46 insertions(+)
> > > create mode 100644 gcc/testsuite/gcc.dg/pr109393.c
> > >
> > > diff --git a/gcc/match.pd b/gcc/match.pd
> > > index d401e7503e6..13c828ba70d 100644
> > > --- a/gcc/match.pd
> > > +++ b/gcc/match.pd
> > > @@ -3650,6 +3650,36 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> > > (plus (convert @0) (op @2 (convert @1))))))
> > > #endif
> > >
> > > +/* ((T)(A + CST1)) * CST2 + CST3
> > > + -> ((T)(A) * CST2) + ((T)CST1 * CST2 + CST3)
> > > + Where (A + CST1) doesn't need to have a single use. */
> > > +#if GIMPLE
> > > + (for op (plus minus)
> > > + (simplify
> > > + (plus (mult (convert:s (op @0 INTEGER_CST@1)) INTEGER_CST@2)
> > > INTEGER_CST@3)
> > > + (if (TREE_CODE (TREE_TYPE (@0)) == INTEGER_TYPE
> > > + && TREE_CODE (type) == INTEGER_TYPE
> > > + && TYPE_PRECISION (type) > TYPE_PRECISION (TREE_TYPE (@0))
> > > + && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))
> > > + && !TYPE_OVERFLOW_SANITIZED (TREE_TYPE (@0))
> > > + && TYPE_OVERFLOW_WRAPS (type))
> > > + (op (mult @2 (convert @0)) (plus (mult @2 (convert @1)) @3)))))
> > > +#endif
> > > +
> > > +/* ((T)(A + CST1)) * CST2 -> ((T)(A) * CST2) + ((T)CST1 * CST2) */
> > > +#if GIMPLE
> > > + (for op (plus minus)
> > > + (simplify
> > > + (mult (convert:s (op:s @0 INTEGER_CST@1)) INTEGER_CST@2)
> > > + (if (TREE_CODE (TREE_TYPE (@0)) == INTEGER_TYPE
> >
> > Please use INTEGRAL_TYPE_P
> >
> > > + && TREE_CODE (type) == INTEGER_TYPE
> > > + && TYPE_PRECISION (type) > TYPE_PRECISION (TREE_TYPE (@0))
> > > + && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))
> > > + && !TYPE_OVERFLOW_SANITIZED (TREE_TYPE (@0))
> > > + && TYPE_OVERFLOW_WRAPS (type))
> > > + (op (mult @2 (convert @0)) (mult @2 (convert @1))))))
> >
> > (mult @2 (convert @0)) is non-canonical for no good reason if @0
> > isn't constant - constant should be 2nd, please swap operands here.
> >
> > > +#endif
> >
> > The first pattern is an extension of the second, why's the first
> > necessary at all? The add of CST3 is unchanged (OK, you seem to
> > associate here, but that's again a different thing).
> >
> > I'd say the 2nd pattern is OK with the above changes but the first
> > looks redundant.
> >
> Hi Richard,
>
> Thanks for the comments, I'll fix these.
>
> The difference is that the second uses op:s while the first uses just op.
> In the second case if A + CST1 has other uses expanding the pattern
> may not be a good idea but in the first case it always is because we
> know + CST1 * CST2 will merge with + CST3.
I see. But that pattern misses a :s on the multiplication result then,
no? Local pattern-matching isn't the best vehicle to handle multi-use
cases, introducing context dependent canonicalizations can lead to
SCEV analysis no longer matching up for related accesses and then
data dependence analysis failing. It's been a trade-off here.
Richard.
>
> Thanks,
> Manolis
>
> > Thanks,
> > Richard.
> >
> > > +
> > > /* (T)(A) +- (T)(B) -> (T)(A +- B) only when (A +- B) could be simplified
> > > to a simple value. */
> > > (for op (plus minus)
> > > diff --git a/gcc/testsuite/gcc.dg/pr109393.c
> > > b/gcc/testsuite/gcc.dg/pr109393.c
> > > new file mode 100644
> > > index 00000000000..e9051273672
> > > --- /dev/null
> > > +++ b/gcc/testsuite/gcc.dg/pr109393.c
> > > @@ -0,0 +1,16 @@
> > > +/* PR tree-optimization/109393 */
> > > +/* { dg-do compile } */
> > > +/* { dg-options "-O2 -fdump-tree-optimized" } */
> > > +/* { dg-final { scan-tree-dump-times "return 1;" 2 "optimized" } } */
> > > +
> > > +int foo(int *a, int j)
> > > +{
> > > + int k = j - 1;
> > > + return a[j - 1] == a[k];
> > > +}
> > > +
> > > +int bar(int *a, int j)
> > > +{
> > > + int k = j - 1;
> > > + return (&a[j + 1] - 2) == &a[k];
> > > +}
> > >
> >
> > --
> > Richard Biener <rguent...@suse.de>
> > SUSE Software Solutions Germany GmbH,
> > Frankenstrasse 146, 90461 Nuernberg, Germany;
> > GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
