Re: [PATCH 1/3] PHIOPT: Add diamond bb form to factor_out_conditional_conversion

[Richard Biener <richard.guenther@gmail.com>]

On Mon, May 8, 2023 at 9:09 AM Andrew Pinski <pinskia@gmail.com> wrote:
>
> On Sun, May 7, 2023 at 11:53 PM Richard Biener via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > On Mon, May 8, 2023 at 8:51 AM Richard Biener
> > <richard.guenther@gmail.com> wrote:
> > >
> > > On Sun, May 7, 2023 at 7:05 AM Andrew Pinski via Gcc-patches
> > > <gcc-patches@gcc.gnu.org> wrote:
> > > >
> > > > So the function factor_out_conditional_conversion already supports
> > > > diamond shaped bb forms, just need to be called for such a thing.
> > > >
> > > > harden-cond-comp.c needed to be changed as we would optimize out the
> > > > conversion now and that causes the compare hardening not needing to
> > > > split the block which it was testing. So change it such that there
> > > > would be no chance of optimization.
> > > >
> > > > Also add two testcases that showed the improvement. PR 103771 is
> > > > solved in ifconvert also for the vectorizer but now it is solved
> > > > in a general sense.
> > > >
> > > > OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.
> > >
> > > So what does it do for the diamond case?  Factor out _common_ operations
> > > only or also promote conditionally executed operations to unconditionally
> > > executed?  Apart from cost considerations (with the looping patch any number
> > > might end up promoted) there's correctness issues for negation (signed overflow)
> > > and for non-call exceptions there's possible external throws.
> > >
> > > If it only factors out common operations the patch is OK (but the testcases
> > > suggest that's not the case).
>
> Right because the testcases are also testing if there was another
> phiopt that happened in the same pass rather than having to wait
> again.
>
> >> Otherwise we should add a limit as to how many
> > > ops we hoist (maybe one for the diamond case?).  Thinking over it the same
> > > issue of course exists for the non-diamond case?
> >
> > Oh, and we hoist the stuff without being sure the final PHI will be if-converted
> > in the end, correct?  Can we somehow improve on that, aka tentatively
> > convert if-convert the PHI first?
>
> Yes it hoists the stuff without any checks on if the final PHI will
> have happened.
> We have the same issue with hoisting adjacent loads.
> I will think of a way to handle to maybe handle this better and even
> undoing it; I will have to get back to you the details though.
> And yes that was already an existing issue with the non-diamond case
> of this; there was some extra checks added to prevent moving some
> casts from being moved out of the PHI due to extra zero extends with
> respect of __builtin_popcount/clz, etc.

OK.  Would be nice if you can somehow handle this.  Even adding a
static limit to the number of hoisted ops isn't good - multiple phiopt
passes will simply run up to N times such limit :/  Undoing might be
"easiest", best would be separating analysis and transform phase
somehow.

Let's go with the patch as-is in the meantime.

Thanks,
Richard.

> Thanks,
> Andrew
>
>
>
> >
> > Richard.
> >
> > > Richard.
> > >
> > > >         PR tree-optimization/49959
> > > >         PR tree-optimization/103771
> > > >
> > > > gcc/ChangeLog:
> > > >
> > > >         * tree-ssa-phiopt.cc (pass_phiopt::execute): Support
> > > >         Diamond shapped bb form for factor_out_conditional_conversion.
> > > >
> > > > gcc/testsuite/ChangeLog:
> > > >
> > > >         * c-c++-common/torture/harden-cond-comp.c: Change testcase
> > > >         slightly to avoid the new phiopt optimization.
> > > >         * gcc.dg/tree-ssa/abs-2.c: New test.
> > > >         * gcc.dg/tree-ssa/pr103771.c: New test.
> > > > ---
> > > >  .../c-c++-common/torture/harden-cond-comp.c   |  6 +++---
> > > >  gcc/testsuite/gcc.dg/tree-ssa/abs-2.c         | 20 +++++++++++++++++++
> > > >  gcc/testsuite/gcc.dg/tree-ssa/pr103771.c      | 18 +++++++++++++++++
> > > >  gcc/tree-ssa-phiopt.cc                        |  2 +-
> > > >  4 files changed, 42 insertions(+), 4 deletions(-)
> > > >  create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/abs-2.c
> > > >  create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/pr103771.c
> > > >
> > > > diff --git a/gcc/testsuite/c-c++-common/torture/harden-cond-comp.c b/gcc/testsuite/c-c++-common/torture/harden-cond-comp.c
> > > > index 5aad890a1d3..dcf364ee993 100644
> > > > --- a/gcc/testsuite/c-c++-common/torture/harden-cond-comp.c
> > > > +++ b/gcc/testsuite/c-c++-common/torture/harden-cond-comp.c
> > > > @@ -1,11 +1,11 @@
> > > >  /* { dg-do compile } */
> > > >  /* { dg-options "-fharden-conditional-branches -fharden-compares -fdump-tree-hardcbr -fdump-tree-hardcmp -ffat-lto-objects" } */
> > > >
> > > > -int f(int i, int j) {
> > > > +int f(int i, int j, int k, int l) {
> > > >    if (i == 0)
> > > > -    return j != 0;
> > > > +    return (j != 0) + l;
> > > >    else
> > > > -    return i * j != 0;
> > > > +    return (i * j != 0) * k;
> > > >  }
> > > >
> > > >  /* { dg-final { scan-tree-dump-times "Splitting edge" 2 "hardcbr" } } */
> > > > diff --git a/gcc/testsuite/gcc.dg/tree-ssa/abs-2.c b/gcc/testsuite/gcc.dg/tree-ssa/abs-2.c
> > > > new file mode 100644
> > > > index 00000000000..328b1802541
> > > > --- /dev/null
> > > > +++ b/gcc/testsuite/gcc.dg/tree-ssa/abs-2.c
> > > > @@ -0,0 +1,20 @@
> > > > +/* PR tree-optimization/49959 */
> > > > +/* { dg-do compile } */
> > > > +/* { dg-options "-O -fdump-tree-phiopt1-details" } */
> > > > +
> > > > +#define ABS(X)    (((X)>0)?(X):-(X))
> > > > +unsigned long
> > > > +test_abs(int *cur)
> > > > +{
> > > > +  unsigned long sad = 0;
> > > > +  if (cur[0] > 0)
> > > > +    sad = cur[0];
> > > > +  else
> > > > +    sad = -cur[0];
> > > > +  return sad;
> > > > +}
> > > > +
> > > > +/* We should figure out that test_abs has an ABS_EXPR in it. */
> > > > +/* { dg-final { scan-tree-dump " = ABS_EXPR" "phiopt1"} } */
> > > > +/* { dg-final { scan-tree-dump-times "changed to factor conversion out from" 1 "phiopt1"} } */
> > > > +
> > > > diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr103771.c b/gcc/testsuite/gcc.dg/tree-ssa/pr103771.c
> > > > new file mode 100644
> > > > index 00000000000..97c9db846cb
> > > > --- /dev/null
> > > > +++ b/gcc/testsuite/gcc.dg/tree-ssa/pr103771.c
> > > > @@ -0,0 +1,18 @@
> > > > +/* { dg-do compile } */
> > > > +/* { dg-options "-O3 -fdump-tree-phiopt1-details" } */
> > > > +/* { dg-final { scan-tree-dump-times "changed to factor conversion out from COND_EXPR." 1 "phiopt1" } } */
> > > > +
> > > > +typedef unsigned char uint8_t;
> > > > +
> > > > +static uint8_t x264_clip_uint8 (int x)
> > > > +{
> > > > +  return x & (~255) ? (-x) >> 31 : x;
> > > > +}
> > > > +
> > > > +void
> > > > +mc_weight (uint8_t* __restrict dst, uint8_t* __restrict src,
> > > > +          int i_width,int i_scale)
> > > > +{
> > > > +  for(int x = 0; x < i_width; x++)
> > > > +    dst[x] = x264_clip_uint8 (src[x] * i_scale);
> > > > +}
> > > > diff --git a/gcc/tree-ssa-phiopt.cc b/gcc/tree-ssa-phiopt.cc
> > > > index f14b7e8b7e6..41fea78dc8d 100644
> > > > --- a/gcc/tree-ssa-phiopt.cc
> > > > +++ b/gcc/tree-ssa-phiopt.cc
> > > > @@ -4087,7 +4087,7 @@ pass_phiopt::execute (function *)
> > > >
> > > >        gphi *newphi;
> > > >        if (single_pred_p (bb1)
> > > > -         && !diamond_p
> > > > +         && EDGE_COUNT (merge->preds) == 2
> > > >           && (newphi = factor_out_conditional_conversion (e1, e2, phi,
> > > >                                                           arg0, arg1,
> > > >                                                           cond_stmt)))
> > > > --
> > > > 2.31.1
> > > >