[Bug tree-optimization/64493] [4.8/4.9/5 Regression] ICE at -O3 on x86_64-linux-gnu

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64493

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
I tried to make a runtime testcase that tests for proper vectorization but
all vectorized code ends up being dead.  Nevertheless the following fixes
the ICE:

Index: gcc/tree-vect-loop.c
===================================================================
--- gcc/tree-vect-loop.c        (revision 219520)
+++ gcc/tree-vect-loop.c        (working copy)
@@ -4580,7 +4580,10 @@ vect_finalize_reduction:
                            && !STMT_VINFO_LIVE_P (exit_phi_vinfo))
                           || double_reduc);

-              STMT_VINFO_VEC_STMT (exit_phi_vinfo) = epilog_stmt;
+             if (double_reduc)
+               STMT_VINFO_VEC_STMT (exit_phi_vinfo) = inner_phi;
+             else
+               STMT_VINFO_VEC_STMT (exit_phi_vinfo) = epilog_stmt;
               if (!double_reduc
                   || STMT_VINFO_DEF_TYPE (exit_phi_vinfo)
                       != vect_double_reduction_def)

and the generated code looks reasonable.  I think we don't really support
the case where the inner reduction result is used in the outer loop (but
in !relevant stmts which end up dead).

Well.  Not 100% sure.