On Thu, Jan 8, 2026 at 11:20 AM Robin Dapp <[email protected]> wrote:
>
> Hi,
>
> RVV's vectors can get very large with LMUL8. In the PR we have
> 256-element char vectors which get permuted. For permuting them
> we use a mask vectype that is deduced from the element type
> without checking if the permute indices fit this type.
> That leads to an invalid permute mask which gets optimized away.
>
> This patch punts if the permute index might overflow the mask type.
>
> Bootstrapped and regtested on x86, power10, and aarch64.
> Regtested on riscv64.
>
> Regards
> Robin
>
> PR tree-optimization/123414
>
> gcc/ChangeLog:
>
> * tree-ssa-forwprop.cc (simplify_vector_constructor):
> Give up if the permute index might overflow the mask type.
Ok with one minor change listed below:
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/riscv/rvv/autovec/pr123414.c: New test.
> ---
> .../gcc.target/riscv/rvv/autovec/pr123414.c | 31 +++++++++++++++++++
> gcc/tree-ssa-forwprop.cc | 13 ++++++--
> 2 files changed, 41 insertions(+), 3 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/autovec/pr123414.c
>
> diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr123414.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr123414.c
> new file mode 100644
> index 00000000000..a28ce23f058
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr123414.c
> @@ -0,0 +1,31 @@
> +/* { dg-do compile } */
> +/* { dg-options "-march=rv64gcv_zvl256b -mrvv-max-lmul=m8 -O3 -fsigned-char
> -fno-strict-aliasing -fwrapv -fdump-tree-optimized -std=gnu99" } */
> +
> +signed char a=2;
> +long long b;
> +long c = 93;
> +int e[1][9];
> +
> +void
> +g (long cc, int ee[][9])
> +{
> + for (int i = 0; i < 4; i++)
> + for (int j = 0; j < 5; j++)
> + for (unsigned k = 0; k < 9; k++)
> + {
> + a *= cc;
> + for (int l = 0; l < 6; l += (ee[k] <= 0) + 2)
> + ;
> + }
> +}
> +
> +int main() {
> + g( c, e);
> + b = (int)a;
> + if (b != 34)
> + __builtin_abort ();
> +}
> +
> +/* We should have four vector constructors that must not get optimized away.
> + */
> +/* { dg-final { scan-tree-dump-times
> "\[a-zA-Z_\]\[a-zA-Z0-9_\]+.=.\\\{\[a-zA-Z0-9._\]+, 1, 1, 1, 1, 1, 1, 1, 1,
> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1," 4
> "optimized" } } */
> diff --git a/gcc/tree-ssa-forwprop.cc b/gcc/tree-ssa-forwprop.cc
> index d00f7b9004e..9435e4dc409 100644
> --- a/gcc/tree-ssa-forwprop.cc
> +++ b/gcc/tree-ssa-forwprop.cc
> @@ -4177,9 +4177,16 @@ simplify_vector_constructor (gimple_stmt_iterator *gsi)
> machine_mode vmode = TYPE_MODE (perm_type);
> if (!can_vec_perm_const_p (vmode, vmode, indices))
> return false;
> - mask_type
> - = build_vector_type (build_nonstandard_integer_type (elem_size, 1),
> - refnelts);
> + /* With an ELEM_SIZEd integer we can shuffle at most
> + (1 << ELEM_SIZE) / 2 - 1 elements. For a more fine-grained
> + check we could take the maximum of the element values
> + but for now this is sufficient. */
> + if (refnelts < (HOST_WIDE_INT_1U << (elem_size - 1)))
> + mask_type
> + = build_vector_type (build_nonstandard_integer_type (elem_size, 1),
> + refnelts);
> + else
> + return false;
Just do:
if (refnelts >= (HOST_WIDE_INT_1U << (elem_size - 1)))
return false;
And then don't touch the whitespace on the mask_type line.
Thanks,
Andrew Pinski
> if (GET_MODE_CLASS (TYPE_MODE (mask_type)) != MODE_VECTOR_INT
> || maybe_ne (GET_MODE_SIZE (TYPE_MODE (mask_type)),
> GET_MODE_SIZE (TYPE_MODE (perm_type))))
> --
> 2.52.0
>
