On Fri, Nov 17, 2023 at 8:20 PM Robin Dapp <rdapp....@gmail.com> wrote:
>
> > No, you shouldn't place _7 != 0 inside the .COND_ADD but instead
> > have an extra pattern stmt producing that so
> >
> > patt_8 = _7 != 0;
> > patt_9 = .COND_ADD (patt_8, ...);
> >
> > that's probably still not enough, but I always quickly forget how
> > bool patterns work ... basically a comparison like patt_8 = _7 != 0
> > vectorizes to a mask (aka vector boolean) while any "data" uses
> > of bools are replaced by mask ? 1 : 0; - there's a complication for
> > bool data producing loads which is why we need to insert the
> > "fake" compares to produce a mask. IIRC.
>
> I already had call handling to vect_recog_bool_pattern in working
> shape when I realized that vect_recog_mask_conversion_pattern already
> handles most of what I need. The difference is that it doesn't do
> patt_8 = _7 != 0
> but rather
> patt_8 = (<signed-boolean:1>) _7;
>
> It works equally well and most of the code can be reused.
>
> The attached was bootstrapped and regtested on x86 and aarch64
> and regtested on riscv.
OK.
Thanks,
Richard.
> Regards
> Robin
>
> Subject: [PATCH] vect: Add bool pattern handling for COND_OPs.
>
> In order to handle masks properly for conditional operations this patch
> teaches vect_recog_mask_conversion_pattern to also handle conditional
> operations. Now we convert e.g.
>
> _mask = *_6;
> _ifc123 = COND_OP (_mask, ...);
>
> into
> _mask = *_6;
> patt200 = (<signed-boolean:1>) _mask;
> patt201 = COND_OP (patt200, ...);
>
> This way the mask will be properly recognized as boolean mask and the
> correct vector mask will be generated.
>
> gcc/ChangeLog:
>
> PR middle-end/112406
>
> * tree-vect-patterns.cc (build_mask_conversion):
> (vect_convert_mask_for_vectype):
>
> gcc/testsuite/ChangeLog:
>
> * gfortran.dg/pr112406.f90: New test.
> ---
> gcc/testsuite/gfortran.dg/pr112406.f90 | 21 +++++++++++++++++++++
> gcc/tree-vect-patterns.cc | 26 ++++++++++++++++++--------
> 2 files changed, 39 insertions(+), 8 deletions(-)
> create mode 100644 gcc/testsuite/gfortran.dg/pr112406.f90
>
> diff --git a/gcc/testsuite/gfortran.dg/pr112406.f90
> b/gcc/testsuite/gfortran.dg/pr112406.f90
> new file mode 100644
> index 00000000000..27e96df7e26
> --- /dev/null
> +++ b/gcc/testsuite/gfortran.dg/pr112406.f90
> @@ -0,0 +1,21 @@
> +! { dg-do compile { target { aarch64-*-* || riscv*-*-* } } }
> +! { dg-options "-Ofast -w -fprofile-generate" }
> +! { dg-additional-options "-march=rv64gcv -mabi=lp64d" { target riscv*-*-* }
> }
> +! { dg-additional-options "-march=armv8-a+sve" { target aarch64-*-* } }
> +
> +module brute_force
> + integer, parameter :: r=9
> + integer sudoku1(1, r)
> + contains
> +subroutine brute
> +integer l(r), u(r)
> + where(sudoku1(1, :) /= 1)
> + l = 1
> + u = 1
> + end where
> +do i1 = 1, u(1)
> + do
> + end do
> + end do
> +end
> +end
> diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> index 7debe7f0731..696b70b76a8 100644
> --- a/gcc/tree-vect-patterns.cc
> +++ b/gcc/tree-vect-patterns.cc
> @@ -5830,7 +5830,8 @@ vect_recog_mask_conversion_pattern (vec_info *vinfo,
> tree rhs1_op0 = NULL_TREE, rhs1_op1 = NULL_TREE;
> tree rhs1_op0_type = NULL_TREE, rhs1_op1_type = NULL_TREE;
>
> - /* Check for MASK_LOAD ans MASK_STORE calls requiring mask conversion. */
> + /* Check for MASK_LOAD and MASK_STORE as well as COND_OP calls requiring
> mask
> + conversion. */
> if (is_gimple_call (last_stmt)
> && gimple_call_internal_p (last_stmt))
> {
> @@ -5842,6 +5843,7 @@ vect_recog_mask_conversion_pattern (vec_info *vinfo,
> return NULL;
>
> bool store_p = internal_store_fn_p (ifn);
> + bool load_p = internal_store_fn_p (ifn);
> if (store_p)
> {
> int rhs_index = internal_fn_stored_value_index (ifn);
> @@ -5856,15 +5858,21 @@ vect_recog_mask_conversion_pattern (vec_info *vinfo,
> vectype1 = get_vectype_for_scalar_type (vinfo, TREE_TYPE (lhs));
> }
>
> + if (!vectype1)
> + return NULL;
> +
> tree mask_arg = gimple_call_arg (last_stmt, mask_argno);
> tree mask_arg_type = integer_type_for_mask (mask_arg, vinfo);
> - if (!mask_arg_type)
> - return NULL;
> - vectype2 = get_mask_type_for_scalar_type (vinfo, mask_arg_type);
> + if (mask_arg_type)
> + {
> + vectype2 = get_mask_type_for_scalar_type (vinfo, mask_arg_type);
>
> - if (!vectype1 || !vectype2
> - || known_eq (TYPE_VECTOR_SUBPARTS (vectype1),
> - TYPE_VECTOR_SUBPARTS (vectype2)))
> + if (!vectype2
> + || known_eq (TYPE_VECTOR_SUBPARTS (vectype1),
> + TYPE_VECTOR_SUBPARTS (vectype2)))
> + return NULL;
> + }
> + else if (store_p || load_p)
> return NULL;
>
> tmp = build_mask_conversion (vinfo, mask_arg, vectype1, stmt_vinfo);
> @@ -5883,7 +5891,9 @@ vect_recog_mask_conversion_pattern (vec_info *vinfo,
> lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
> gimple_call_set_lhs (pattern_stmt, lhs);
> }
> - gimple_call_set_nothrow (pattern_stmt, true);
> +
> + if (load_p || store_p)
> + gimple_call_set_nothrow (pattern_stmt, true);
>
> pattern_stmt_info = vinfo->add_stmt (pattern_stmt);
> if (STMT_VINFO_DATA_REF (stmt_vinfo))
> --
> 2.41.0
>
