https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128695
Bug ID: 128695
Summary: Missed early mode pruning in vect_analyze_loop when
chaining multiple epilogues (vectorizer)
Product: gcc
Version: 15.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: jin_kang66 at 163 dot com
Target Milestone: ---
Location: gcc/tree-vect-loop.cc (function vect_analyze_loop)
Problem Description:
--------------------
In vect_analyze_loop, when analyzing vector modes for epilogue loops, there is
an early-filtering check intended to skip candidate vector modes that cannot
possibly yield a valid epilogue vectorization factor (VF):
bool supports_partial_vectors = partial_vectors_supported_p ()
&& param_vect_partial_vector_usage != 0;
poly_uint64 first_vinfo_vf = LOOP_VINFO_VECT_FACTOR (first_loop_vinfo);
loop_vec_info orig_loop_vinfo = first_loop_vinfo;
do
{
while (1)
{
/* If the target does not support partial vectors we can shorten the
number of modes to analyze for the epilogue as we know we can't
pick
a mode that would lead to a VF at least as big as the
FIRST_VINFO_VF. */
if (!supports_partial_vectors
&& maybe_ge (cached_vf_per_mode[mode_i], first_vinfo_vf))
{
mode_i++;
if (mode_i == vector_modes.length ())
break;
continue;
}
When epilogue chaining was added via the outer do { ... } while (1) loop,
orig_loop_vinfo is updated to advance to the next epilogue (orig_loop_vinfo =
orig_loop_vinfo->epilogue_vinfo).
However, the check still compares against first_vinfo_vf (the primary vector
loop's VF), rather than the immediate parent loop's VF (LOOP_VINFO_VECT_FACTOR
(orig_loop_vinfo)).
Impact:
For subsequent chained epilogues (e.g. searching for epilogue 2 after
epilogue 1):
- Suppose Main Loop has VF = 16 (AVX-512) -> first_vinfo_vf = 16.
- Epilogue 1 is chosen with VF = 8 (AVX2).
- When searching for Epilogue 2, a valid mode must have VF < 8 (e.g. SSE, VF
= 4).
- If another mode in vector_modes has a cached VF of 8: maybe_ge (8, 16)
evaluates to false, so the mode is NOT pruned early.
- GCC invokes vect_analyze_loop_1, running unnecessary data reference,
alignment, dependence, and SLP analyses, only to have vect_analyze_loop_2
reject it downstream via:
if (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)
&& maybe_ge (LOOP_VINFO_VECT_FACTOR (loop_vinfo),
LOOP_VINFO_VECT_FACTOR (orig_loop_vinfo)))
return opt_result::failure_at (...);
While this does not cause a correctness issue (as vect_analyze_loop_2 acts as a
safety net), it misses an early pruning opportunity and leads to redundant
compilation overhead.
Suggested Fix:
Inside the do { ... } while (1) loop, the check should compare against the
current parent loop's VF:
diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
--- a/gcc/tree-vect-loop.cc
+++ b/gcc/tree-vect-loop.cc
@@ -...,... +...,... @@ vect_analyze_loop (class loop *loop, gimple
*loop_vectorized_call,
loop_vec_info orig_loop_vinfo = first_loop_vinfo;
do
{
+ poly_uint64 parent_vf = LOOP_VINFO_VECT_FACTOR (orig_loop_vinfo);
while (1)
{
/* If the target does not support partial vectors we can shorten the
number of modes to analyze for the epilogue as we know we can't
pick
a mode that would lead to a VF at least as big as the
- FIRST_VINFO_VF. */
+ PARENT_VF. */
if (!supports_partial_vectors
- && maybe_ge (cached_vf_per_mode[mode_i], first_vinfo_vf))
+ && maybe_ge (cached_vf_per_mode[mode_i], parent_vf))
{
mode_i++;
if (mode_i == vector_modes.length ())