Apologies for the prolonged silence Richard, it is a bit of an obscure topic,
and I was unsure I'd be able to handle any complications in a timely manner.
I'm ready to revisit it now, please see below.
On Mon, 17 Jan 2022, Richard Biener wrote:
> On Fri, Jan 14, 2022 at 7:21 PM Alexander Monakov <amona...@ispras.ru> wrote:
> >
> > A returns_twice call may have associated abnormal edges that correspond
> > to the "second return" from the call. If the call is duplicated, the
> > copies of those edges also need to be abnormal, but e.g. tracer does not
> > enforce that. Just prohibit the (unlikely to be useful) duplication.
>
> The general CFG copying routines properly duplicate those edges, no?
No (in fact you say so in the next paragraph). In general I think they cannot,
abnormal edges are a special case, so it should be the responsibility of the
caller.
> Tracer uses duplicate_block so it should also get copies of all successor
> edges of that block. It also only traces along normal edges. What it might
> miss is abnormal incoming edges - is that what you are referring to?
Yes (I think its entire point is to build a "trace" of duplicated blocks that
does not have incoming edges in the middle, abnormal or not).
> That would be a thing we don't handle in duplicate_block on its own but
> that callers are expected to do (though I don't see copy_bbs doing that
> either). I wonder if we can trigger this issue for some testcase?
Oh yes (in fact my desire to find a testcase delayed this quite a bit).
When compiling the following testcase with -O2 -ftracer:
__attribute__((returns_twice))
int rtwice_a(int), rtwice_b(int);
int f(int *x)
{
volatile unsigned k, i = (*x);
for (k = 1; (i = rtwice_a(i)) * k; k = 2);
for (; (i = rtwice_b(i)) * k; k = 4);
return k;
}
tracer manages to eliminate the ABNORMAL_DISPATCHER block completely, so
the possibility of transferring control back to rtwice_a from rtwice_b
is no longer modeled in the IR. I could spend some time "upgrading" this
to an end-to-end miscompilation, but I hope you agree this is quite broken
already.
> The thing to check would be incoming abnormal edges in
> can_duplicate_block_p, not (only) returns twice functions?
Unfortunately not, abnormal edges are also used for computed gotos, which are
less magic than returns_twice edges and should not block tracer I think.
This implies patch 1/3 [1] unnecessary blocks sinking to computed goto targets.
[1] https://gcc.gnu.org/pipermail/gcc-patches/2022-January/588498.html
How would you like to proceed here? Is my initial patch ok?
Alexander
>
> Richard.
>
> > gcc/ChangeLog:
> >
> > * tree-cfg.c (gimple_can_duplicate_bb_p): Reject blocks with
> > calls that may return twice.
> > ---
> > gcc/tree-cfg.c | 7 +++++--
> > 1 file changed, 5 insertions(+), 2 deletions(-)
> >
> > diff --git a/gcc/tree-cfg.c b/gcc/tree-cfg.c
> > index b7fe313b7..a99f1acb4 100644
> > --- a/gcc/tree-cfg.c
> > +++ b/gcc/tree-cfg.c
> > @@ -6304,12 +6304,15 @@ gimple_can_duplicate_bb_p (const_basic_block bb)
> > {
> > gimple *g = gsi_stmt (gsi);
> >
> > - /* An IFN_GOMP_SIMT_ENTER_ALLOC/IFN_GOMP_SIMT_EXIT call must be
> > + /* Prohibit duplication of returns_twice calls, otherwise associated
> > + abnormal edges also need to be duplicated properly.
> > + An IFN_GOMP_SIMT_ENTER_ALLOC/IFN_GOMP_SIMT_EXIT call must be
> > duplicated as part of its group, or not at all.
> > The IFN_GOMP_SIMT_VOTE_ANY and IFN_GOMP_SIMT_XCHG_* are part of
> > such a
> > group, so the same holds there. */
> > if (is_gimple_call (g)
> > - && (gimple_call_internal_p (g, IFN_GOMP_SIMT_ENTER_ALLOC)
> > + && (gimple_call_flags (g) & ECF_RETURNS_TWICE
> > + || gimple_call_internal_p (g, IFN_GOMP_SIMT_ENTER_ALLOC)
> > || gimple_call_internal_p (g, IFN_GOMP_SIMT_EXIT)
> > || gimple_call_internal_p (g, IFN_GOMP_SIMT_VOTE_ANY)
> > || gimple_call_internal_p (g, IFN_GOMP_SIMT_XCHG_BFLY)
> > --
> > 2.33.1
> >
>
