[Bug tree-optimization/98117] [8/9/10/11 Regression] wrong code with "-O3 -fno-tree-scev-cprop" since r8-1163-g7078979b291419f3

rguenth at gcc dot gnu.org gcc-bugzilla@gcc.gnu.org
Mon Dec 7 09:27:42 GMT 2020


https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98117

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rsandifo at gcc dot gnu.org

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
In fact there is alrady

      /* Create: niters >> log2(vf) */
      /* If it's known that niters == number of latch executions + 1 doesn't
         overflow, we can generate niters >> log2(vf); otherwise we generate
         (niters - vf) >> log2(vf) + 1 by using the fact that we know ratio
         will be at least one.  */

so we know about this "defect".  The computation method translates easily
to the range case (unconditionally).  Unfortunately for some 'degenerate'
cases where we arrive with const_vf == 1 we'll compute [1, 0] "ranges"
this way which ICEs in niter compute.

The question is whether we can construct cases we miscompile with such
large niter and VF == 1 (full SLP).  As noted when I added niter_m1 all
uses of 'niter' would have to go away :/

For now I'm testing the "ugly"

diff --git a/gcc/tree-vect-loop-manip.c b/gcc/tree-vect-loop-manip.c
index 36179188f6d..36f218e7f6e 100644
--- a/gcc/tree-vect-loop-manip.c
+++ b/gcc/tree-vect-loop-manip.c
@@ -2034,13 +2034,21 @@ vect_gen_vector_loop_niters (loop_vec_info loop_vinfo,
tree niters,
       niters_vector = force_gimple_operand (niters_vector, &stmts, true, var);
       gsi_insert_seq_on_edge_immediate (pe, stmts);
       /* Peeling algorithm guarantees that vector loop bound is at least ONE,
-        we set range information to make niters analyzer's life easier.  */
+        we set range information to make niters analyzer's life easier.
+        Note the number of latch iteration value can be TYPE_MAX_VALUE so
+        we have to represent the vector niter TYPE_MAX_VALUE + 1 >> log_vf. 
*/
       if (stmts != NULL && log_vf)
        set_range_info (niters_vector, VR_RANGE,
-                       wi::to_wide (build_int_cst (type, 1)),
-                       wi::to_wide (fold_build2 (RSHIFT_EXPR, type,
-                                                 TYPE_MAX_VALUE (type),
-                                                 log_vf)));
+                       wi::one (TYPE_PRECISION (type)),
+                       /* ???  Avoid creating [1, 0].  */
+                       const_vf == 1
+                       ? wi::max_value (TYPE_PRECISION (type),
+                                        TYPE_SIGN (type))
+                       : (wi::rshift (wi::max_value (TYPE_PRECISION (type),
+                                                     TYPE_SIGN (type))
+                                      - (const_vf - 1),
+                                      exact_log2 (const_vf), TYPE_SIGN (type))
+                          + 1));
     }
   *niters_vector_ptr = niters_vector;
   *step_vector_ptr = step_vector;


More information about the Gcc-bugs mailing list