On Tue, Mar 25, 2025 at 04:44:45PM +0100, Jakub Jelinek wrote:
> On Tue, Mar 25, 2025 at 08:33:41AM -0700, Andi Kleen wrote:
> > > 2025-03-25 Jakub Jelinek <ja...@redhat.com>
> > > Andi Kleen <a...@gcc.gnu.org>
> > >
> > > PR gcov-profile/118442
> > > * profile.cc (branch_prob): Ignore EDGE_FAKE edges from musttail calls
> > > to EXIT.
> > >
> > > * c-c++-common/pr118442.c: New test.
> > >
> > > --- gcc/profile.cc.jj 2025-01-02 11:23:16.458517673 +0100
> > > +++ gcc/profile.cc 2025-03-25 09:57:21.860398601 +0100
> > > @@ -1340,6 +1340,20 @@ branch_prob (bool thunk)
> > > EDGE_INFO (e)->ignore = 1;
> > > ignored_edges++;
> > > }
> > > + /* Ignore fake edges after musttail calls. */
> > > + if ((e->flags & EDGE_FAKE)
> > > + && e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
> > > + {
> > > + gimple_stmt_iterator gsi = gsi_last_bb (e->src);
> >
> > At least the musttail pass allows some statements after the call, like
> > labels
> > and debug information. Not sure if it matters.
>
> I think it shouldn't.
> gimple_flow_call_edges_add splits basic blocks with calls after those calls
> and adds the EDGE_FAKE edges to EXIT, so the last stmt at the end of
> the bb from which the edge goes should be the call.
In particular
/* Note that the following may create a new basic block
and renumber the existing basic blocks. */
if (stmt != last_stmt)
{
e = split_block (bb, stmt);
if (e)
blocks_split++;
}
e = make_edge (bb, EXIT_BLOCK_PTR_FOR_FN (cfun), EDGE_FAKE);
Anyway, the patch has been bootstrapped/regtested on x86_64-linux and
i686-linux successfully. Ok for trunk?
Jakub
