Re: [PATCH] [GCC] AutoFDO: Walk predecessor edges for PHI arguments

On Tue, Aug 4, 2026 at 2:39 AM Kugan Vivekanandarajah <
[email protected]> wrote:

> PHI arguments are indexed by incoming edges.  function_instance::match
> walked successor edges and could fail gimple_phi_arg_def_from_edge's
> destination assertion during afdo_offline.
>
> gcc/ChangeLog:
>
>         * auto-profile.cc (function_instance::match): Walk predecessor
> edges
>         for PHI arguments.
>

This is OK.  It is really strange this does not trigger checking failure
I see that
inline location_t
gimple_phi_arg_location_from_edge (gphi *phi, edge e)
{
 return gimple_phi_arg (phi, e->dest_idx)->locus;

Does not check that PHI BB and e->dest match (that looks like omission that
would be useful to fix), but still we should eventually hit out of range
access which is checked in gimple_phi_arg.

Honza

>
> Regression tested on aarch64-linux-gnu with no new regressions
>
> Signed-off-by: Kugan Vivekanandarajah <[email protected]>
> ---
>  gcc/auto-profile.cc | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/gcc/auto-profile.cc b/gcc/auto-profile.cc
> index f4562d7f2f98..60f6b1ba480b 100644
> --- a/gcc/auto-profile.cc
> +++ b/gcc/auto-profile.cc
> @@ -1764,7 +1764,8 @@ function_instance::match (cgraph_node *node,
>           gcc_assert (!info);
>           dump_stmt (phi, info, NULL, stack);
>           counts.add (info);
> -         for (edge e : bb->succs)
> +         /* PHI arguments are indexed by incoming (predecessor) edges.  */
> +         for (edge e : bb->preds)
>             {
>               location_t phi_loc
>                 = gimple_phi_arg_location_from_edge (phi, e);
> --
> 2.34.1
>
>