[PATCH] Move negative stride bias out of dr_misalignment

[Richard Biener]

On Tue, 26 Oct 2021, Bernhard Reutner-Fischer wrote:

> On 26 October 2021 11:19:44 CEST, Richard Biener via Gcc-patches <gcc-patches@gcc.gnu.org> wrote:
> 
> 
>  
> >@@ -2010,6 +2010,7 @@ get_negative_load_store_type (vec_info *vinfo,
> >       if (dump_enabled_p ())
> > 	dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> > 			 "negative step but alignment required.\n");
> >+      *poffset = 0;
> >       return VMAT_ELEMENTWISE;
> >       *poffset = 0;
> >     }
> 
> I think you cannot really diagnose these, they would trigger a lot, in each early return, no?
> Or would we see that there are unreachable, non artificial stmts in the same block after a return?
> Somebody could experiment with diagnosing each and every DCEd stmt (!artificial, nondebug) but I would expect that hell breaks loose..
> cheers,

[I fixed the issue before pushing this patch btw]

I agree in general, where I would diagnose this is when we build the
CFG I would diagnose unreachable blocks - the above does not have
any path to the second *poffset store.  Like with the prototype patch
below we warn for

int *p;
int foo()
{
  return 0;
  *p = 0;
}
> ./cc1 -quiet t.c
t.c: In function 'foo':
t.c:5:3: warning: statement is not reachable
    5 |   *p = 0;
      |   ^~

diff --git a/gcc/tree-cfg.c b/gcc/tree-cfg.c
index b3a27bcd17c..2de661ba35d 100644
--- a/gcc/tree-cfg.c
+++ b/gcc/tree-cfg.c
@@ -374,6 +374,22 @@ execute_build_cfg (void)
       fprintf (dump_file, "Scope blocks:\n");
       dump_scope_blocks (dump_file, dump_flags);
     }
+
+  find_unreachable_blocks ();
+  basic_block bb;
+  FOR_EACH_BB_FN (bb, cfun)
+    if (!(bb->flags & BB_REACHABLE))
+      for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p 
(gsi);
+          gsi_next (&gsi))
+       {
+         if (gimple_location (gsi_stmt (gsi)) != UNKNOWN_LOCATION)
+           warning_at (gimple_location (gsi_stmt (gsi)), 0,
+                       "statement is not reachable");
+         /* ???  Mark blocks reachable from here.  And even better make
+            sure to process entries to unreachable regions first.  */
+         break;
+       }
+
   cleanup_tree_cfg ();
 
   bb_to_omp_idx.release ();