Hi,
since peeling and version for alignment for VLA modes was introduced
(r16-3065-geee51f9a4b6) we have been seeing a lot of test suite failures
like
internal compiler error: in apply_scale, at profile-count.h:1187
This is because vect_gen_prolog_loop_niters sets the prolog bound to -1
in case align_in_elems is a non-constant poly_int.
bound - 1 is later used to scale the loop profile in scale_loop_profile
so we try to calculate with an assumed -2 iterations.
This patch sets bound to 0 instead so scale_loop_profile gets passed -1
and won't do anything then.
Regtested on rv64gcv_zvl512b and aarch64 via qemu.
Regards
Robin
gcc/ChangeLog:
* tree-vect-loop-manip.cc (PROLOG_BOUND_UNKNOWN): Define.
(vect_gen_prolog_loop_niters): Use.
(vect_do_peeling): Use.
---
gcc/tree-vect-loop-manip.cc | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
index 566308f4fe5..578ee82571f 100644
--- a/gcc/tree-vect-loop-manip.cc
+++ b/gcc/tree-vect-loop-manip.cc
@@ -2505,6 +2505,8 @@ get_misalign_in_elems (gimple **seq, loop_vec_info
loop_vinfo)
the number of elements that fit in the vector. Therefore, instead of VF we
use TYPE_VECTOR_SUBPARTS. */
+#define PROLOG_BOUND_UNKNOWN (0)
+
static tree
vect_gen_prolog_loop_niters (loop_vec_info loop_vinfo,
basic_block bb, int *bound)
@@ -2559,7 +2561,7 @@ vect_gen_prolog_loop_niters (loop_vec_info loop_vinfo,
if (align_in_elems.is_constant (&align_in_elems_c))
*bound = align_in_elems_c - 1;
else
- *bound = -1;
+ *bound = PROLOG_BOUND_UNKNOWN;
}
if (dump_enabled_p ())
@@ -3257,7 +3259,7 @@ vect_do_peeling (loop_vec_info loop_vinfo, tree niters,
tree nitersm1,
/* Generate the number of iterations for the prolog loop. We do this here
so that we can also get the upper bound on the number of iterations. */
tree niters_prolog;
- int bound_prolog = 0;
+ int bound_prolog = PROLOG_BOUND_UNKNOWN;
if (prolog_peeling)
{
niters_prolog = vect_gen_prolog_loop_niters (loop_vinfo, anchor,
--
2.50.0
