[Bug tree-optimization/106476] New: ICE generating FOLD_EXTRACT_LAST

[ams at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106476

            Bug ID: 106476
           Summary: ICE generating FOLD_EXTRACT_LAST
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: major
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ams at gcc dot gnu.org
                CC: rguenther at suse dot de
  Target Milestone: ---
            Target: amdgcn-amdhsa

Commit 8f4d9c1deda "amdgcn: 64-bit not" exposed an ICE in tree-vect_stmts.cc
when compiling gcc.dg/torture/pr67470.c at -O2 for amdgcn. The newly
implemented op is not the problem, but it allows this optimization (and many
others) to proceed, and the error is no longer hidden.

amdgcn has masked vectors and fold_extract_last, which leads to a code path
through tree-vect-stmts.cc that has

   vec_then_clause = vec_oprnds2[i];
   if (reduction_type != EXTRACT_LAST_REDUCTION)
     vec_else_clause = vec_oprnds3[i];

and then

   /* Instead of doing ~x ? y : z do x ? z : y.  */
   vec_compare = new_temp;
   std::swap (vec_then_clause, vec_else_clause);

and finally

   new_stmt = gimple_build_call_internal
       (IFN_FOLD_EXTRACT_LAST, 3, else_clause, vec_compare,
        vec_then_clause);

in which vec_then_clause remains set to NULL_TREE.

The dump shows

   e_lsm.16_32 = .FOLD_EXTRACT_LAST (e_lsm.16_8, _70, );

(note the last field is missing.)

I can fix the ICE if I add "else vec_else_clause = integer_zero_node", but I'm
not sure that is the correct logical solution.

(CC Richi who touched this code last)