On Mon, May 9, 2022 at 7:11 AM liuhongt via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> Here's adjused patch.
> Ok for trunk?
>
> Optimize
>
> _4 = VEC_PERM_EXPR <_1, _1, { 4, 5, 6, 7, 4, 5, 6, 7 }>;
> _5 = BIT_FIELD_REF <_4, 128, 0>;
>
> to
>
> _5 = BIT_FIELD_REF <_1, 128, 128>;
>
> gcc/ChangeLog:
>
> PR tree-optimization/102583
> * tree-ssa-forwprop.cc (simplify_bitfield_ref): Extended to a
> contiguous stride in the VEC_PERM_EXPR.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/pr102583.c: New test.
> * gcc.target/i386/pr92645-2.c: Adjust testcase.
> * gcc.target/i386/pr92645-3.c: Ditto.
> ---
> gcc/testsuite/gcc.target/i386/pr102583.c | 30 ++++++++
> gcc/testsuite/gcc.target/i386/pr92645-2.c | 4 +-
> gcc/testsuite/gcc.target/i386/pr92645-3.c | 4 +-
> gcc/tree-ssa-forwprop.cc | 89 ++++++++++++++++-------
> 4 files changed, 96 insertions(+), 31 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/i386/pr102583.c
>
> diff --git a/gcc/testsuite/gcc.target/i386/pr102583.c
> b/gcc/testsuite/gcc.target/i386/pr102583.c
> new file mode 100644
> index 00000000000..4ef2f296d0c
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr102583.c
> @@ -0,0 +1,30 @@
> +/* { dg-do compile } */
> +/* { dg-options "-mavx512f -O2" } */
> +/* { dg-final { scan-assembler-times {(?n)vcvtdq2ps[ \t]+32\(%.*%ymm} 1 } }
> */
> +/* { dg-final { scan-assembler-times {(?n)vcvtdq2ps[ \t]+16\(%.*%xmm} 1 } }
> */
> +/* { dg-final { scan-assembler-times {(?n)vmovq[ \t]+16\(%.*%xmm} 1 { target
> { ! ia32 } } } } */
> +/* { dg-final { scan-assembler-not {(?n)vpermd[ \t]+.*%zmm} } } */
> +
> +typedef int v16si __attribute__((vector_size(64)));
> +typedef float v8sf __attribute__((vector_size(32)));
> +typedef float v4sf __attribute__((vector_size(16)));
> +typedef float v2sf __attribute__((vector_size(8)));
> +
> +v8sf part (v16si *srcp)
> +{
> + v16si src = *srcp;
> + return (v8sf) { (float)src[8], (float) src[9], (float)src[10],
> (float)src[11],
> + (float)src[12], (float)src[13], (float)src[14], (float)src[15] };
> +}
> +
> +v4sf part1 (v16si *srcp)
> +{
> + v16si src = *srcp;
> + return (v4sf) { (float)src[4], (float)src[5], (float)src[6], (float)src[7]
> };
> +}
> +
> +v2sf part2 (v16si *srcp)
> +{
> + v16si src = *srcp;
> + return (v2sf) { (float)src[4], (float)src[5] };
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr92645-2.c
> b/gcc/testsuite/gcc.target/i386/pr92645-2.c
> index d34ed3aa8e5..f0608de938a 100644
> --- a/gcc/testsuite/gcc.target/i386/pr92645-2.c
> +++ b/gcc/testsuite/gcc.target/i386/pr92645-2.c
> @@ -29,6 +29,6 @@ void odd (v2si *dst, v4si *srcp)
> }
>
> /* { dg-final { scan-tree-dump-times "BIT_FIELD_REF" 4 "cddce1" } } */
> -/* { dg-final { scan-tree-dump-times "VEC_PERM_EXPR" 3 "cddce1" } } */
> +/* { dg-final { scan-tree-dump-times "VEC_PERM_EXPR" 3 "cddce1" { xfail
> *-*-* } } } */
> /* Ideally highpart extraction would elide the permutation as well. */
> -/* { dg-final { scan-tree-dump-times "VEC_PERM_EXPR" 2 "cddce1" { xfail
> *-*-* } } } */
> +/* { dg-final { scan-tree-dump-times "VEC_PERM_EXPR" 2 "cddce1" } } */
> diff --git a/gcc/testsuite/gcc.target/i386/pr92645-3.c
> b/gcc/testsuite/gcc.target/i386/pr92645-3.c
> index 9c08c9fb632..691011195c9 100644
> --- a/gcc/testsuite/gcc.target/i386/pr92645-3.c
> +++ b/gcc/testsuite/gcc.target/i386/pr92645-3.c
> @@ -32,6 +32,6 @@ void odd (v4sf *dst, v8si *srcp)
> /* Four conversions, on the smaller vector type, to not convert excess
> elements. */
> /* { dg-final { scan-tree-dump-times " = \\\(vector\\\(4\\\) float\\\)" 4
> "cddce1" } } */
> -/* { dg-final { scan-tree-dump-times "VEC_PERM_EXPR" 3 "cddce1" } } */
> +/* { dg-final { scan-tree-dump-times "VEC_PERM_EXPR" 3 "cddce1" { xfail
> *-*-* } } } */
> /* Ideally highpart extraction would elide the VEC_PERM_EXPR as well. */
> -/* { dg-final { scan-tree-dump-times "VEC_PERM_EXPR" 2 "cddce1" { xfail
> *-*-* } } } */
> +/* { dg-final { scan-tree-dump-times "VEC_PERM_EXPR" 2 "cddce1" } } */
> diff --git a/gcc/tree-ssa-forwprop.cc b/gcc/tree-ssa-forwprop.cc
> index 484491fa1c5..f91f738895d 100644
> --- a/gcc/tree-ssa-forwprop.cc
> +++ b/gcc/tree-ssa-forwprop.cc
> @@ -2334,8 +2334,10 @@ simplify_bitfield_ref (gimple_stmt_iterator *gsi)
> gimple *stmt = gsi_stmt (*gsi);
> gimple *def_stmt;
> tree op, op0, op1;
> - tree elem_type;
> - unsigned idx, size;
> + tree elem_type, type;
> + tree p, m, tem;
> + unsigned HOST_WIDE_INT nelts;
> + unsigned idx, size, elem_size;
> enum tree_code code;
>
> op = gimple_assign_rhs1 (stmt);
> @@ -2353,42 +2355,75 @@ simplify_bitfield_ref (gimple_stmt_iterator *gsi)
> op1 = TREE_OPERAND (op, 1);
> code = gimple_assign_rhs_code (def_stmt);
> elem_type = TREE_TYPE (TREE_TYPE (op0));
> - if (TREE_TYPE (op) != elem_type)
> + type = TREE_TYPE (op);
> + /* Also hanlde vector type.
> + .i.e.
> + _7 = VEC_PERM_EXPR <_1, _1, { 2, 3, 2, 3 }>;
> + _11 = BIT_FIELD_REF <_7, 64, 0>;
> +
> + to
> +
> + _11 = BIT_FIELD_REF <_1, 64, 64>. */
> + if (type != elem_type
> + && (!VECTOR_TYPE_P (type) || TREE_TYPE (type) != elem_type))
> return false;
I'm not sure we actually need this check. I think we could handle
a __int128_t extract from the permuted vector as well, no?
>
> - size = TREE_INT_CST_LOW (TYPE_SIZE (elem_type));
> + elem_size = size = TREE_INT_CST_LOW (TYPE_SIZE (type));
That's not going to work with VLA vectors, I think you need to make
elem_size/size poly_uint64 and use tree_to_poly_uint64 (TYPE_SIZE (type))
here.
> if (maybe_ne (bit_field_size (op), size))
> return false;
>
> - if (code == VEC_PERM_EXPR
> - && constant_multiple_p (bit_field_offset (op), size, &idx))
> + if (code != VEC_PERM_EXPR
> + || !constant_multiple_p (bit_field_offset (op), size, &idx))
That might be a bit restrictive? I think we want elem_size here
with elem_size picked from the vector type of the VEC_PERM_EXPR
operand.
> + return false;
> +
> + m = gimple_assign_rhs3 (def_stmt);
> + if (TREE_CODE (m) != VECTOR_CST
> + || !VECTOR_CST_NELTS (m).is_constant (&nelts))
> + return false;
> +
> + /* One element. */
> + if (type == elem_type)
I think this should work on the sizes, so
if (known_eq (size, elem_size))
...
that makes sure we do not end up with nelts_op == 1 below and also
handle the case where the BIT_FIELD_REF would do int<->float
punning.
> + idx = TREE_INT_CST_LOW (VECTOR_CST_ELT (m, idx));
> + else
> {
> - tree p, m, tem;
> - unsigned HOST_WIDE_INT nelts;
> - m = gimple_assign_rhs3 (def_stmt);
> - if (TREE_CODE (m) != VECTOR_CST
> - || !VECTOR_CST_NELTS (m).is_constant (&nelts))
> + elem_size = TREE_INT_CST_LOW (TYPE_SIZE (elem_type));
> + unsigned nelts_op;
> + if (!constant_multiple_p (bit_field_size (op), elem_size, &nelts_op))
bit_field_size (op) == size? I think we also want to handle
power-of-two nelts_op
only (I can't think of how we get a non-power-of-two, but checking
pow2p_hwi (nelts_op)
woudl be nice if only for documentation purposes)
> return false;
> - idx = TREE_INT_CST_LOW (VECTOR_CST_ELT (m, idx));
> - idx %= 2 * nelts;
> - if (idx < nelts)
> - {
> - p = gimple_assign_rhs1 (def_stmt);
> - }
> - else
> + unsigned start = TREE_INT_CST_LOW (vector_cst_elt (m, idx));
> + unsigned end = TREE_INT_CST_LOW (vector_cst_elt (m, idx + nelts_op -
> 1));
> + /* Be in the same vector. */
> + if ((start < nelts) != (end < nelts))
> + return false;
> + for (unsigned HOST_WIDE_INT i = 1; i != nelts_op; i++)
> {
> - p = gimple_assign_rhs2 (def_stmt);
> - idx -= nelts;
> + /* Continuous area. */
> + if (TREE_INT_CST_LOW (vector_cst_elt (m, idx + i)) - 1
> + != TREE_INT_CST_LOW (vector_cst_elt (m, idx + i - 1)))
> + return false;
> }
> - tem = build3 (BIT_FIELD_REF, TREE_TYPE (op),
> - unshare_expr (p), op1, bitsize_int (idx * size));
> - gimple_assign_set_rhs1 (stmt, tem);
> - fold_stmt (gsi);
> - update_stmt (gsi_stmt (*gsi));
> - return true;
> + /* Alignment not worse than before. */
> + unsigned dest_align = TREE_INT_CST_LOW (TYPE_SIZE (type));
> + if (start * elem_size % dest_align)
> + return false;
So I think we want to have start * elem_size be a multiple of size
(the check I said above is too restrictive on the _original_ bit-field-ref but
we want to have for the target bit-field-ref). I think that's what you check
here but see above for the use of poly-ints.
> + idx = start;
> }
>
> - return false;
> + idx %= 2 * nelts;
that's odd - why do you need to clamp idx?
> + if (idx < nelts)
> + p = gimple_assign_rhs1 (def_stmt);
> + else
> + {
> + p = gimple_assign_rhs2 (def_stmt);
> + idx -= nelts;
> + }
> +
> + tem = build3 (BIT_FIELD_REF, TREE_TYPE (op),
> + unshare_expr (p), op1, bitsize_int (idx * elem_size));
You don't need to unshare_expr 'p' here.
Otherwise looks OK.
Thanks,
Richard.
> + gimple_assign_set_rhs1 (stmt, tem);
> + fold_stmt (gsi);
> + update_stmt (gsi_stmt (*gsi));
> + return true;
> }
>
> /* Determine whether applying the 2 permutations (mask1 then mask2)
> --
> 2.18.1
>
