For AMD GCN, the instructions available for loading/storing vectors are
always scatter/gather operations (i.e. there are separate addresses for
each vector lane), so the current heuristic to avoid gather/scatter
operations with too many elements in get_group_load_store_type is
counterproductive. Avoiding such operations in that function can
subsequently lead to a missed vectorization opportunity whereby later
analyses in the vectorizer try to use a very wide array type which is
not available on this target, and thus it bails out.
The attached patch adds a target hook to override the "single_element_p"
heuristic in the function as a target hook, and activates it for GCN. This
allows much better code to be generated for affected loops.
Tested with offloading to AMD GCN. I will apply to the og10 branch
shortly.
Julian
2021-01-13 Julian Brown <jul...@codesourcery.com>
gcc/
* doc/tm.texi.in (TARGET_VECTORIZE_PREFER_GATHER_SCATTER): Add
documentation hook.
* doc/tm.texi: Regenerate.
* target.def (prefer_gather_scatter): Add target hook under vectorizer.
* tree-vect-stmts.c (get_group_load_store_type): Optionally prefer
gather/scatter instructions to scalar/elementwise fallback.
* config/gcn/gcn.c (TARGET_VECTORIZE_PREFER_GATHER_SCATTER): Define
hook.
---
gcc/config/gcn/gcn.c | 2 ++
gcc/doc/tm.texi | 5 +++++
gcc/doc/tm.texi.in | 2 ++
gcc/target.def | 8 ++++++++
gcc/tree-vect-stmts.c | 9 +++++++--
5 files changed, 24 insertions(+), 2 deletions(-)
diff --git a/gcc/config/gcn/gcn.c b/gcc/config/gcn/gcn.c
index ee9f00558305..ea88b5e91244 100644
--- a/gcc/config/gcn/gcn.c
+++ b/gcc/config/gcn/gcn.c
@@ -6501,6 +6501,8 @@ gcn_dwarf_register_span (rtx rtl)
gcn_vector_alignment_reachable
#undef TARGET_VECTOR_MODE_SUPPORTED_P
#define TARGET_VECTOR_MODE_SUPPORTED_P gcn_vector_mode_supported_p
+#undef TARGET_VECTORIZE_PREFER_GATHER_SCATTER
+#define TARGET_VECTORIZE_PREFER_GATHER_SCATTER true
struct gcc_target targetm = TARGET_INITIALIZER;
diff --git a/gcc/doc/tm.texi b/gcc/doc/tm.texi
index 581b7b51eeb0..bd0b2eea477a 100644
--- a/gcc/doc/tm.texi
+++ b/gcc/doc/tm.texi
@@ -6122,6 +6122,11 @@ The default is @code{NULL_TREE} which means to not
vectorize scatter
stores.
@end deftypefn
+@deftypevr {Target Hook} bool TARGET_VECTORIZE_PREFER_GATHER_SCATTER
+This hook is set to TRUE if gather loads or scatter stores are cheaper on
+this target than a sequence of elementwise loads or stores.
+@end deftypevr
+
@deftypefn {Target Hook} int TARGET_SIMD_CLONE_COMPUTE_VECSIZE_AND_SIMDLEN
(struct cgraph_node *@var{}, struct cgraph_simd_clone *@var{}, @var{tree},
@var{int})
This hook should set @var{vecsize_mangle}, @var{vecsize_int},
@var{vecsize_float}
fields in @var{simd_clone} structure pointed by @var{clone_info} argument and
also
diff --git a/gcc/doc/tm.texi.in b/gcc/doc/tm.texi.in
index afa19d4ac63c..c0883e5da82c 100644
--- a/gcc/doc/tm.texi.in
+++ b/gcc/doc/tm.texi.in
@@ -4195,6 +4195,8 @@ address; but often a machine-dependent strategy can
generate better code.
@hook TARGET_VECTORIZE_BUILTIN_SCATTER
+@hook TARGET_VECTORIZE_PREFER_GATHER_SCATTER
+
@hook TARGET_SIMD_CLONE_COMPUTE_VECSIZE_AND_SIMDLEN
@hook TARGET_SIMD_CLONE_ADJUST
diff --git a/gcc/target.def b/gcc/target.def
index 00421f3a6acd..0b34ab5c3d52 100644
--- a/gcc/target.def
+++ b/gcc/target.def
@@ -2027,6 +2027,14 @@ all zeros. GCC can then try to branch around the
instruction instead.",
(unsigned ifn),
default_empty_mask_is_expensive)
+/* Prefer gather/scatter loads/stores to e.g. elementwise accesses if\n\
+we cannot use a contiguous access. */
+DEFHOOKPOD
+(prefer_gather_scatter,
+ "This hook is set to TRUE if gather loads or scatter stores are cheaper on\n\
+this target than a sequence of elementwise loads or stores.",
+ bool, false)
+
/* Target builtin that implements vector gather operation. */
DEFHOOK
(builtin_gather,
diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c
index 9ace345fc5e2..e117d3d16afc 100644
--- a/gcc/tree-vect-stmts.c
+++ b/gcc/tree-vect-stmts.c
@@ -2444,9 +2444,14 @@ get_group_load_store_type (stmt_vec_info stmt_info, tree
vectype, bool slp,
it probably isn't a win to use separate strided accesses based
on nearby locations. Or, even if it's a win over scalar code,
it might not be a win over vectorizing at a lower VF, if that
- allows us to use contiguous accesses. */
+ allows us to use contiguous accesses.
+
+ On some targets (e.g. AMD GCN), always use gather/scatter accesses
+ here since those are the only types of vector loads/stores available,
+ and the fallback case of using elementwise accesses is very
+ inefficient. */
if (*memory_access_type == VMAT_ELEMENTWISE
- && single_element_p
+ && (targetm.vectorize.prefer_gather_scatter || single_element_p)
&& loop_vinfo
&& vect_use_strided_gather_scatters_p (stmt_info, loop_vinfo,
masked_p, gs_info))
--
2.29.2
