On Wed, May 13, 2026 at 8:48 AM Andrew Pinski
<[email protected]> wrote:
>
> On Tue, May 12, 2026 at 11:30 PM Richard Biener
> <[email protected]> wrote:
> >
> > On Tue, May 12, 2026 at 11:58 PM Andrew Pinski
> > <[email protected]> wrote:
> > >
> > > Canonicalize `(type) (a CMP b)` to `(a CMP b) ? (type) true : (type)
> > > false` for
> > > floating point types. fold-const already does this so it makes sense to
> > > do it on
> > > the gimple level too. This also improves vectorization because a
> > > conversion
> > > from a comparison to a float does not need to implemented and also if it
> > > is
> > > implemented then selection of 1.0/0.0 is easier than the conversion would
> > > be.
> > >
> > > Bootstrapped and tested on x86_64-linux-gnu with no regressions.
> > >
> > > PR tree-optimization/115571
> > >
> > > gcc/ChangeLog:
> > >
> > > * match.pd (`(float)(a CMP b)`): New pattern.
> > >
> > > gcc/testsuite/ChangeLog:
> > >
> > > * gcc.dg/tree-ssa/float-conv-1.c: New test.
> > >
> > > Signed-off-by: Andrew Pinski <[email protected]>
> > > ---
> > > gcc/match.pd | 6 ++++++
> > > gcc/testsuite/gcc.dg/tree-ssa/float-conv-1.c | 16 ++++++++++++++++
> > > 2 files changed, 22 insertions(+)
> > > create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/float-conv-1.c
> > >
> > > diff --git a/gcc/match.pd b/gcc/match.pd
> > > index 8dcf8c318cf..6f1c53e5d7b 100644
> > > --- a/gcc/match.pd
> > > +++ b/gcc/match.pd
> > > @@ -2733,6 +2733,12 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> > > )
> > > )
> > >
> > > +/* Canonicalize `(type) (a CMP b)` to `(a CMP b) ? (type) true : (type)
> > > false`
> > > + for floating point types. */
> > > +(simplify
> > > + (float (convert:s (tcc_comparison@0 @1 @2)))
> >
> > So we'll never end up with (float (tcc_comparison ...))?
>
> I thought we converted bool to int first but I was wrong.
> Testcase for that:
> ```
> float f(float a)
> {
> bool t = a < 0.5;
> return t;
> }
> ```
Note I think we should require a mode-precision integer
source for FLOAT_EXPR (and a mode-precision integer
for FIX_TRUNC_EXPR) and declare bool -> float invalid
GIMPLE. Does RTL expansion even "deal" with this
correctly? You can try various "fancy" GIMPLE testcases for this ...
>
> >
> > Patch is OK, I was just curious.
>
> I will add `?` to convert and retest it with an added testcase and
> push it tomorrow since the `?` addition is an obvious improvement.
>
> Thanks,
> Andrew
>
> >
> > Richard.
> >
> > > + (cond @0 { build_one_cst (type); } { build_zero_cst (type); } ))
> > > +
> > > /* For integral types with undefined overflow and C != 0 fold
> > > x * C EQ/NE y * C into x EQ/NE y. */
> > > (for cmp (eq ne)
> > > diff --git a/gcc/testsuite/gcc.dg/tree-ssa/float-conv-1.c
> > > b/gcc/testsuite/gcc.dg/tree-ssa/float-conv-1.c
> > > new file mode 100644
> > > index 00000000000..ffb07c4b94d
> > > --- /dev/null
> > > +++ b/gcc/testsuite/gcc.dg/tree-ssa/float-conv-1.c
> > > @@ -0,0 +1,16 @@
> > > +/* PR tree-optimization/115571 */
> > > +/* { dg-do compile } */
> > > +/* { dg-options "-O1 -fdump-tree-optimized-raw" } */
> > > +
> > > +float f(float a)
> > > +{
> > > + int t = a < 0.5;
> > > + return t;
> > > +}
> > > +
> > > +float f0(float a)
> > > +{
> > > + return (short)(a < 0.5);
> > > +}
> > > +
> > > +/* { dg-final { scan-tree-dump-times "gimple_cond |gimple_assign
> > > <cond_expr" 2 "optimized" } } */
> > > --
> > > 2.43.0
> > >
