From: juewang <[email protected]>
The unsigned narrow-clip pattern currently builds MAX_EXPR and
IFN_SAT_TRUNC using the promoted input type. This can make the
vectorizer widen an input only to narrow it again.
Look through a single-use promotion when the sign is unchanged and the
target supports SAT_TRUNC directly from the original type. Build the
MAX_EXPR from that input so that the operation retains the narrower
vector element type.
gcc/ChangeLog:
* tree-vect-patterns.cc (vect_recog_sat_trunc_pattern): Use an
unpromoted input when possible.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/sat/vec_sat_u_trunc-promotion-1.c:
New test.
---
.../autovec/sat/vec_sat_u_trunc-promotion-1.c | 28 ++++++++++++
gcc/tree-vect-patterns.cc | 45 ++++++++++++++++---
2 files changed, 66 insertions(+), 7 deletions(-)
create mode 100644
gcc/testsuite/gcc.target/riscv/rvv/autovec/sat/vec_sat_u_trunc-promotion-1.c
diff --git
a/gcc/testsuite/gcc.target/riscv/rvv/autovec/sat/vec_sat_u_trunc-promotion-1.c
b/gcc/testsuite/gcc.target/riscv/rvv/autovec/sat/vec_sat_u_trunc-promotion-1.c
new file mode 100644
index 00000000000..d8f40cb8e88
--- /dev/null
+++
b/gcc/testsuite/gcc.target/riscv/rvv/autovec/sat/vec_sat_u_trunc-promotion-1.c
@@ -0,0 +1,28 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -mrvv-max-lmul=m1
-fno-vect-cost-model" } */
+
+#include <stdint-gcc.h>
+
+static inline uint8_t
+clip_uint8 (int x)
+{
+ return x & ~255 ? (-x) >> 31 : x;
+}
+
+__attribute__ ((noipa)) void
+hpel_clip (uint8_t *__restrict dst, const uint8_t *__restrict src, int n)
+{
+ for (int i = 0; i < n; ++i)
+ {
+ int value = (src[i - 2] + src[i + 3]
+ - 5 * (src[i - 1] + src[i + 2])
+ + 20 * (src[i] + src[i + 1]));
+ dst[i] = clip_uint8 ((value + 16) >> 5);
+ }
+}
+
+/* { dg-final { scan-assembler-times {vnclipu\.wi} 1 } } */
+/* { dg-final { scan-assembler {vsetvli\t[^\n]+,e8,mf2,ta,ma} } } */
+/* { dg-final { scan-assembler {vsetvli\t[^\n]+,e16,m1,ta,ma} } } */
+/* { dg-final { scan-assembler-not {vsext\.vf2} } } */
+/* { dg-final { scan-assembler-not {e32,m1} } } */
diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
index b921ae94848..4865a0ded90 100644
--- a/gcc/tree-vect-patterns.cc
+++ b/gcc/tree-vect-patterns.cc
@@ -4822,20 +4822,51 @@ vect_recog_sat_trunc_pattern (vec_info *vinfo,
stmt_vec_info stmt_vinfo,
tree lhs = gimple_assign_lhs (last_stmt);
tree otype = TREE_TYPE (lhs);
- if ((gimple_unsigned_integer_narrow_clip (lhs, ops, NULL))
- && type_has_mode_precision_p (otype))
+ if (gimple_unsigned_integer_narrow_clip (lhs, ops, NULL)
+ && type_has_mode_precision_p (otype))
{
- tree itype = TREE_TYPE (ops[0]);
- tree v_itype = get_vectype_for_scalar_type (vinfo, itype);
+ tree input = ops[0];
+ tree itype = TREE_TYPE (input);
tree v_otype = get_vectype_for_scalar_type (vinfo, otype);
internal_fn fn = IFN_SAT_TRUNC;
+ /* Avoid widening an input only to narrow it again. A signed
+ promotion does not affect MAX (X, 0), so use the unpromoted value
+ when the target can narrow it directly to the output type. */
+ vect_unpromoted_value unprom;
+ bool single_use_p = true;
+ if (v_otype != NULL_TREE
+ && vect_look_through_possible_promotion (vinfo, input, &unprom,
+ &single_use_p)
+ && single_use_p
+ && INTEGRAL_TYPE_P (unprom.type)
+ && TYPE_SIGN (unprom.type) == TYPE_SIGN (itype)
+ && type_has_mode_precision_p (unprom.type)
+ && TYPE_PRECISION (otype) < TYPE_PRECISION (unprom.type)
+ && TYPE_PRECISION (unprom.type) < TYPE_PRECISION (itype))
+ {
+ tree v_unprom_type
+ = get_vectype_for_scalar_type (vinfo, unprom.type);
+ if (v_unprom_type != NULL_TREE
+ && direct_internal_fn_supported_p
+ (fn, tree_pair (v_otype, v_unprom_type),
+ OPTIMIZE_FOR_BOTH))
+ {
+ input = unprom.op;
+ itype = unprom.type;
+ }
+ }
+
+ tree v_itype = get_vectype_for_scalar_type (vinfo, itype);
if (v_itype != NULL_TREE && v_otype != NULL_TREE
- && direct_internal_fn_supported_p (fn, tree_pair (v_otype, v_itype),
- OPTIMIZE_FOR_BOTH))
+ && direct_internal_fn_supported_p (fn,
+ tree_pair (v_otype, v_itype),
+ OPTIMIZE_FOR_BOTH))
{
tree temp = vect_recog_temp_ssa_var (itype, NULL);
- gimple * max_stmt = gimple_build_assign (temp, build2 (MAX_EXPR,
itype, build_zero_cst(itype), ops[0]));
+ tree max_op = build2 (MAX_EXPR, itype, build_zero_cst (itype),
+ input);
+ gimple *max_stmt = gimple_build_assign (temp, max_op);
append_pattern_def_seq (vinfo, stmt_vinfo, max_stmt, v_itype);
gcall *call = gimple_build_call_internal (fn, 1, temp);
--
2.34.1
