On Thu, Aug 27, 2026 at 10:10 AM wangjue <[email protected]> wrote:
>
> From: juewang <[email protected]>
>
> Use a single-use unpromoted input when the target can narrow it
> directly to the output type. This avoids widening the input only to
> narrow it again.
>
> Add RISC-V tests for x264-style promoted inputs, saturation boundary
> values, and one-to-three-level unsigned narrowing. Disable vector
> costing where the tests require vectorization independent of the
> default tune.
Can you show the former original GIMPLE and matched patterns and the
resulting optimized pattern sequence?
> gcc/ChangeLog:
>
> * tree-vect-patterns.cc (vect_recog_sat_trunc_pattern): Look through
> promotions.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/riscv/rvv/autovec/pr120378-2.c: Dump optimized GIMPLE.
> * gcc.target/riscv/rvv/autovec/pr120378-3.c: Likewise. Disable the
> vector cost model.
> * gcc.target/riscv/rvv/autovec/pr120378-4.c: Likewise.
> * gcc.target/riscv/rvv/autovec/sat-trunc-promote-1.c: New test.
> * gcc.target/riscv/rvv/autovec/sat-trunc-promote-run-1.c: New test.
> ---
> .../gcc.target/riscv/rvv/autovec/pr120378-2.c | 2 +-
> .../gcc.target/riscv/rvv/autovec/pr120378-3.c | 2 +-
> .../gcc.target/riscv/rvv/autovec/pr120378-4.c | 2 +-
> .../riscv/rvv/autovec/sat-trunc-promote-1.c | 25 +++++++
> .../rvv/autovec/sat-trunc-promote-run-1.c | 45 +++++++++++++
> gcc/tree-vect-patterns.cc | 65 ++++++++++++++-----
> 6 files changed, 122 insertions(+), 19 deletions(-)
> create mode 100644
> gcc/testsuite/gcc.target/riscv/rvv/autovec/sat-trunc-promote-1.c
> create mode 100644
> gcc/testsuite/gcc.target/riscv/rvv/autovec/sat-trunc-promote-run-1.c
>
> diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr120378-2.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr120378-2.c
> index 9a880e0f67b..2208e954242 100644
> --- a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr120378-2.c
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr120378-2.c
> @@ -1,5 +1,5 @@
> /* { dg-do compile } */
> -/* { dg-options "-march=rv64gcv -mabi=lp64d -O3" } */
> +/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -fdump-tree-optimized" } */
>
> #include <stdint.h>
>
> diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr120378-3.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr120378-3.c
> index d6b2ddd3f8d..a71f8bfa19a 100644
> --- a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr120378-3.c
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr120378-3.c
> @@ -1,5 +1,5 @@
> /* { dg-do compile } */
> -/* { dg-options "-march=rv64gcv -mabi=lp64d -O3" } */
> +/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -fno-vect-cost-model
> -fdump-tree-optimized" } */
>
> #include <stdint.h>
>
> diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr120378-4.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr120378-4.c
> index 4657e52f883..bc9a55e8b07 100644
> --- a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr120378-4.c
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr120378-4.c
> @@ -1,5 +1,5 @@
> /* { dg-do compile } */
> -/* { dg-options "-march=rv64gcv -mabi=lp64d -O3" } */
> +/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -fno-vect-cost-model
> -fdump-tree-optimized" } */
>
> #include <stdint.h>
>
> diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/sat-trunc-promote-1.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/sat-trunc-promote-1.c
> new file mode 100644
> index 00000000000..00768de993c
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/sat-trunc-promote-1.c
> @@ -0,0 +1,25 @@
> +/* { dg-do compile { target { rv64 } } } */
> +/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -mrvv-max-lmul=m1
> -fno-vect-cost-model -fdump-tree-optimized" } */
> +
> +#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-tree-dump-times {\.SAT_TRUNC } 1 "optimized" } } */
> +/* { dg-final { scan-assembler-times {\tvnclipu\.wi} 1 } } */
> diff --git
> a/gcc/testsuite/gcc.target/riscv/rvv/autovec/sat-trunc-promote-run-1.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/sat-trunc-promote-run-1.c
> new file mode 100644
> index 00000000000..6249defff38
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/sat-trunc-promote-run-1.c
> @@ -0,0 +1,45 @@
> +/* { dg-do run { target { riscv_v && rv64 } } } */
> +/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -mrvv-max-lmul=m1
> -fno-vect-cost-model" } */
> +
> +#include <stdint-gcc.h>
> +
> +#define N 128
> +
> +static inline uint8_t
> +clip_uint8 (int x)
> +{
> + return x & ~255 ? (-x) >> 31 : x;
> +}
> +
> +__attribute__ ((noipa)) static void
> +clip_loop (uint8_t *__restrict dst, const int16_t *__restrict src, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + {
> + int x = (src[i] >> 8) * 256 + (src[i] & 255);
> + dst[i] = clip_uint8 (x);
> + }
> +}
> +
> +int
> +main (void)
> +{
> + static const int16_t values[] = {
> + INT16_MIN, -1, 0, 255, 256, INT16_MAX
> + };
> + static const uint8_t expected[] = { 0, 0, 0, 255, 255, 255 };
> + int16_t input[N];
> + uint8_t output[N];
> +
> + for (int i = 0; i < N; ++i)
> + input[i] = values[i % 6];
> +
> + clip_loop (output, input, N);
> +
> +#pragma GCC novector
> + for (int i = 0; i < N; ++i)
> + if (output[i] != expected[i % 6])
> + __builtin_abort ();
> +
> + return 0;
> +}
> diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> index 406f81b95c8..c7c5f80c4e9 100644
> --- a/gcc/tree-vect-patterns.cc
> +++ b/gcc/tree-vect-patterns.cc
> @@ -4843,31 +4843,64 @@ vect_recog_sat_trunc_pattern (vec_info *vinfo,
> stmt_vec_info stmt_vinfo,
> 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;
>
> - if (v_itype != NULL_TREE && v_otype != NULL_TREE
> - && direct_internal_fn_supported_p (fn, tree_pair (v_otype, v_itype),
> - OPTIMIZE_FOR_BOTH))
> + if (v_otype != NULL_TREE)
> {
> - 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]));
> - append_pattern_def_seq (vinfo, stmt_vinfo, max_stmt, v_itype);
> + /* 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 (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;
> + }
> + }
>
> - gcall *call = gimple_build_call_internal (fn, 1, temp);
> - tree out_ssa = vect_recog_temp_ssa_var (otype, NULL);
> + tree v_itype = get_vectype_for_scalar_type (vinfo, itype);
> + if (v_itype != NULL_TREE
> + && 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),
> input));
Please do not build a GENERIC MAX_EXPR, instead use the appropriate
gimple_build_assign overload (this also suggests this was LLM generated).
Also I fail to see a check that the generated MAX_EXPR is supported by
the target?
> + append_pattern_def_seq (vinfo, stmt_vinfo, max_stmt, v_itype);
>
> - gimple_call_set_lhs (call, out_ssa);
> - gimple_call_set_nothrow (call, /* nothrow_p */ false);
> - gimple_set_location (call, gimple_location (last_stmt));
> + gcall *call = gimple_build_call_internal (fn, 1, temp);
> + tree out_ssa = vect_recog_temp_ssa_var (otype, NULL);
>
> - *type_out = v_otype;
> + gimple_call_set_lhs (call, out_ssa);
> + gimple_call_set_nothrow (call, false);
> + gimple_set_location (call, gimple_location (last_stmt));
>
> - return call;
> - }
> + *type_out = v_otype;
>
> + return call;
> + }
> + }
> }
>
> if ((gimple_unsigned_integer_sat_trunc (lhs, ops, NULL)
> --
> 2.34.1
>
