[PATCH, v3] Fix PR51513, switch statement with default case containing __builtin_unreachable leads to wild branch

[Peter Bergner]

On 4/27/17 6:57 AM, Bernhard Reutner-Fischer wrote:
> On Wed, Apr 26, 2017 at 10:39:12PM -0500, Peter Bergner wrote:
>> +/* Returns true if the basic block BB has no successors and only contains
>> +   a call to __builtin_unreachable ().  */
> 
> so
>   return EDGE_COUNT (bb->succs) == 0
>     && (gsi = gsi_last_nondebug_bb (bb))
>     && !gsi_end_p (gsi)
>     && gimple_call_builtin_p (gsi_stmt (gsi), BUILT_IN_UNREACHABLE);
[snip]
> which i should better phrase as
>   gsi = gsi_start_nondebug_after_labels_bb (bb);
>   while (!gsi_end_p (gsi)
>     {
>       gimple *stmt = gsi_stmt (gsi);
>       if (gimple_clobber_p (stmt))
> 	gsi_next (&gsi);
>       else
>         return gimple_call_builtin_p (stmt, BUILT_IN_UNREACHABLE);
>     }
>   return false;

I didn't try to rewrite the code too much, I basically just outlined
the code as is.  However, one change I did have to make, was to not
use routines like gsi_after_labels(), etc., since those do not work
during expansion to RTL, due to bb->flags == BB_RTL.  Those off limits
routines include gsi_start_nondebug_after_labels_bb(), since it calls
gsi_after_labels().

Peter