[PATCH 2/2] Fix vect_need_peeling_or_partial_vectors_p wrt peeling for gaps

[Richard Biener]

The code assuming that peeling for gaps can be elided by means of
knowing niters and prologue peeling is wrong.  Peeling for gaps
means we need an epilog.

Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.

        * tree-vect-loop.cc (vect_need_peeling_or_partial_vectors_p):
        When peeling for gaps we always need an epilog.
---
 gcc/tree-vect-loop.cc | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
index cd62c96892b..3787e7bdfb6 100644
--- a/gcc/tree-vect-loop.cc
+++ b/gcc/tree-vect-loop.cc
@@ -946,6 +946,9 @@ vect_need_peeling_or_partial_vectors_p (loop_vec_info 
loop_vinfo)
 {
   unsigned HOST_WIDE_INT const_vf;
 
+  if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo))
+    return true;
+
   loop_vec_info main_loop_vinfo
     = (LOOP_VINFO_EPILOGUE_P (loop_vinfo)
        ? LOOP_VINFO_MAIN_LOOP_INFO (loop_vinfo) : loop_vinfo);
@@ -956,14 +959,11 @@ vect_need_peeling_or_partial_vectors_p (loop_vec_info 
loop_vinfo)
         peeled for reasons other than niters.  */
       unsigned int peel_niter
        = LOOP_VINFO_PEELING_FOR_ALIGNMENT (main_loop_vinfo);
-      if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo))
-       peel_niter += 1;
       return !multiple_p (LOOP_VINFO_INT_NITERS (loop_vinfo) - peel_niter,
                          LOOP_VINFO_VECT_FACTOR (loop_vinfo));
     }
 
   if (!LOOP_VINFO_PEELING_FOR_ALIGNMENT (main_loop_vinfo)
-      && !LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)
       && LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&const_vf))
     {
       /* When the number of iterations is a multiple of the vectorization
-- 
2.51.0