Re: [PATCH] Add extra copy of the ifcombine pass after pre [PR102793]

[Manolis Tsamis <manolis.tsamis@vrull.eu>]

On Fri, Apr 5, 2024 at 3:43 PM Richard Biener
<richard.guenther@gmail.com> wrote:
>
> On Fri, Apr 5, 2024 at 2:28 PM Manolis Tsamis <manolis.tsamis@vrull.eu> wrote:
> >
> > If we consider code like:
> >
> >     if (bar1 == x)
> >       return foo();
> >     if (bar2 != y)
> >       return foo();
> >     return 0;
> >
> > We would like the ifcombine pass to convert this to:
> >
> >     if (bar1 == x || bar2 != y)
> >       return foo();
> >     return 0;
> >
> > The ifcombine pass can handle this transformation but it is ran very early and
> > it misses the opportunity because there are two seperate blocks for foo().
> > The pre pass is good at removing duplicate code and blocks and due to that
> > running ifcombine again after it can increase the number of successful
> > conversions.
> >
> >         PR 102793
> >
> > gcc/ChangeLog:
> >
> >         * common.opt: -ftree-ifcombine option, enabled by default.
> >         * doc/invoke.texi: Document.
> >         * passes.def: Re-run ssa-ifcombine after pre.
> >         * tree-ssa-ifcombine.cc: Make ifcombine cloneable. Add gate function.
> >
> > gcc/testsuite/ChangeLog:
> >
> >         * gcc.dg/tree-ssa/20030922-2.c: Change flag to -fno-tree-ifcombine.
> >         * gcc.dg/uninit-pred-6_c.c: Remove inconsistent check.
> >         * gcc.target/aarch64/pr102793.c: New test.
> >
> > Signed-off-by: Manolis Tsamis <manolis.tsamis@vrull.eu>
> > ---
> >
> >  gcc/common.opt                              |  4 +++
> >  gcc/doc/invoke.texi                         |  5 ++++
> >  gcc/passes.def                              |  1 +
> >  gcc/testsuite/gcc.dg/tree-ssa/20030922-2.c  |  2 +-
> >  gcc/testsuite/gcc.dg/uninit-pred-6_c.c      |  4 ---
> >  gcc/testsuite/gcc.target/aarch64/pr102793.c | 30 +++++++++++++++++++++
> >  gcc/tree-ssa-ifcombine.cc                   |  5 ++++
> >  7 files changed, 46 insertions(+), 5 deletions(-)
> >  create mode 100644 gcc/testsuite/gcc.target/aarch64/pr102793.c
> >
> > diff --git a/gcc/common.opt b/gcc/common.opt
> > index ad348844775..e943202bcf1 100644
> > --- a/gcc/common.opt
> > +++ b/gcc/common.opt
> > @@ -3163,6 +3163,10 @@ ftree-phiprop
> >  Common Var(flag_tree_phiprop) Init(1) Optimization
> >  Enable hoisting loads from conditional pointers.
> >
> > +ftree-ifcombine
>
> Please don't add further -ftree-X flags, 'tree' means nothing
> to users.  -fif-combine would be better.
>
Ok, thanks for pointing out, will do.

> > +Common Var(flag_tree_ifcombine) Init(1) Optimization
> > +Merge some conditional branches to simplify control flow.
> > +
> >  ftree-pre
> >  Common Var(flag_tree_pre) Optimization
> >  Enable SSA-PRE optimization on trees.
> > diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
> > index e2edf7a6c13..8d2ff6b4512 100644
> > --- a/gcc/doc/invoke.texi
> > +++ b/gcc/doc/invoke.texi
> > @@ -13454,6 +13454,11 @@ This flag is enabled by default at @option{-O1} and higher.
> >  Perform hoisting of loads from conditional pointers on trees.  This
> >  pass is enabled by default at @option{-O1} and higher.
> >
> > +@opindex ftree-ifcombine
> > +@item -ftree-ifcombine
> > +Merge some conditional branches to simplify control flow.  This pass
> > +is enabled by default at @option{-O1} and higher.
> > +
> >  @opindex fhoist-adjacent-loads
> >  @item -fhoist-adjacent-loads
> >  Speculatively hoist loads from both branches of an if-then-else if the
> > diff --git a/gcc/passes.def b/gcc/passes.def
> > index 1cbbd413097..1765b476131 100644
> > --- a/gcc/passes.def
> > +++ b/gcc/passes.def
> > @@ -270,6 +270,7 @@ along with GCC; see the file COPYING3.  If not see
> >        NEXT_PASS (pass_lim);
> >        NEXT_PASS (pass_walloca, false);
> >        NEXT_PASS (pass_pre);
> > +      NEXT_PASS (pass_tree_ifcombine);
> >        NEXT_PASS (pass_sink_code, false /* unsplit edges */);
>
> Please move it here, after sinking.
>
My initial placement was there, but I saw that in some cases
(including the testcase), sinking would introduce some blocks that
prohibited the ifcombining.

> >        NEXT_PASS (pass_sancov);
> >        NEXT_PASS (pass_asan);
> > diff --git a/gcc/testsuite/gcc.dg/tree-ssa/20030922-2.c b/gcc/testsuite/gcc.dg/tree-ssa/20030922-2.c
> > index 16c79da9521..66c9f481a2f 100644
> > --- a/gcc/testsuite/gcc.dg/tree-ssa/20030922-2.c
> > +++ b/gcc/testsuite/gcc.dg/tree-ssa/20030922-2.c
> > @@ -1,5 +1,5 @@
> >  /* { dg-do compile } */
> > -/* { dg-options "-O1 -fdump-tree-dom2 -fdisable-tree-ifcombine" } */
> > +/* { dg-options "-O1 -fdump-tree-dom2 -fno-tree-ifcombine" } */
> >
> >  struct rtx_def;
> >  typedef struct rtx_def *rtx;
> > diff --git a/gcc/testsuite/gcc.dg/uninit-pred-6_c.c b/gcc/testsuite/gcc.dg/uninit-pred-6_c.c
> > index f60868dad23..2d8e6501a45 100644
> > --- a/gcc/testsuite/gcc.dg/uninit-pred-6_c.c
> > +++ b/gcc/testsuite/gcc.dg/uninit-pred-6_c.c
> > @@ -20,10 +20,6 @@ int foo (int n, int l, int m, int r)
> >    if ( (n > 10) && l)
> >        blah(v); /* { dg-bogus "uninitialized" "bogus warning" } */
> >
> > -  if (l)
> > -    if (n > 12)
> > -      blah(v); /* { dg-bogus "uninitialized" "bogus warning" } */
> > -
>
> What's "inconsistent" about this check?  I suppose we now diagnose this?
> The appropriate way would be to XFAIL this but I'd like you to explain
> why we now diagnose this (I don't see obvious if-combining opportunities).
>

Maybe inconsistent wasn't the correct word, but what I meant was:
If the condition in the test is merged into if (l && (n > 12)) then
the "bogus warning" appears on master too and the test fails. By
inconsistent I meant that the analysis relies on a particular
arrangement of conditionals that is fragile.

After this change the branch is optimized by ifcombine and the "bogus
warning" appears. We also get better codegen, at least for AArch64.

It felt wrong to just remove this, it didn't occur to me that this can
be made XFAIL instead. I can change it to that if you think it's
justified.

> On a general note you rely on the tail-merging pass which is part of PRE
> and which hasn't seen any love and which isn't very powerful either.  I'm not
> sure it's worth doing if-combining on the whole IL again because of it.
> It might be possible to locally try if-combining from the immediate dominator
> of a merged tail from inside tail-merging itself?
>

I had a feeling that the ifcombine pass is not very expensive and
that's why I opted to just duplicate it (instead of e.g. moving it).
I'm not really familiar with PRE so I made the placement just by
observing the block deduplication.
I don't know the answer to the last part, but if you believe it's
worthwhile I can try it out.

Thanks,
Manolis

> Richard.
>
> >    return 0;
> >  }
> >
> > diff --git a/gcc/testsuite/gcc.target/aarch64/pr102793.c b/gcc/testsuite/gcc.target/aarch64/pr102793.c
> > new file mode 100644
> > index 00000000000..78d48e01637
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/aarch64/pr102793.c
> > @@ -0,0 +1,30 @@
> > +/* { dg-do compile } */
> > +/* { dg-options "-O2" } */
> > +
> > +typedef unsigned long uint64_t;
> > +
> > +int ccmp1(uint64_t* s1, uint64_t* s2, int(*foo)(void))
> > +{
> > +  uint64_t d1, d2, bar;
> > +  d1 = *s1++;
> > +  d2 = *s2++;
> > +  bar = (d1 + d2) & 0xabcd;
> > +  if (bar == 0 || d1 != d2)
> > +    return foo();
> > +  return 0;
> > +}
> > +
> > +int ccmp2(uint64_t* s1, uint64_t* s2, int(*foo)(void))
> > +{
> > +  uint64_t d1, d2, bar;
> > +  d1 = *s1++;
> > +  d2 = *s2++;
> > +  bar = (d1 + d2) & 0xabcd;
> > +  if (bar == 0)
> > +    return foo();
> > +  if (d1 != d2)
> > +    return foo();
> > +  return 0;
> > +}
> > +
> > +/* { dg-final { scan-assembler-times "ccmp\t" 2 } } */
> > \ No newline at end of file
> > diff --git a/gcc/tree-ssa-ifcombine.cc b/gcc/tree-ssa-ifcombine.cc
> > index 6a3bc99190d..0bf9fe8b692 100644
> > --- a/gcc/tree-ssa-ifcombine.cc
> > +++ b/gcc/tree-ssa-ifcombine.cc
> > @@ -838,6 +838,11 @@ public:
> >    {}
> >
> >    /* opt_pass methods: */
> > +  opt_pass * clone () final override
> > +  {
> > +    return new pass_tree_ifcombine (m_ctxt);
> > +  }
> > +  bool gate (function *) final override { return flag_tree_ifcombine; }
> >    unsigned int execute (function *) final override;
> >
> >  }; // class pass_tree_ifcombine
> > --
> > 2.44.0
> >