On Wed, 2 Oct 2019, Jakub Jelinek wrote:
> On Wed, Nov 09, 2016 at 09:14:55AM +0100, Richard Biener wrote:
> > The following implements vectorization of bswap via VEC_PERM_EXPR
> > on the corresponding QImode vector.
> >
> > ARM already has backend handling via the builtin_vectorized_call
> > hook and thus there were already testcases available. It doesn't
> > end up working for vect-bswap16.c because we have a promoted
> > argument to __builtin_bswap16 which confuses vectorization.
>
> Indeed. The following patch handles that in tree-vect-patterns.c.
> If it sees a __builtin_bswap16 with the promoted argument, it checks if we'd
> vectorize the builtin if it didn't have a promoted argument and if yes,
> it just changes it in a pattern_stmt to use an unpromoted argument or casts
> it first to the right type. This works e.g. for the SSE4 case.
> Otherwise, it handles __builtin_bswap16 like a x r<< 8, if that is
> vectorizable, emits a pattern_stmt with x r<< 8, if it isn't, falls back to
> (x << 8) | (x >> 8) if that can be vectorized. The last case matters for
> SSE2.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
OK.
Thanks,
Richard.
> 2019-10-02 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/91940
> * tree-vect-patterns.c: Include tree-vector-builder.h and
> vec-perm-indices.h.
> (vect_recog_rotate_pattern): Also handle __builtin_bswap16, either by
> unpromoting the argument back to uint16_t, or by converting into a
> rotate, or into shifts plus ior.
>
> * gcc.dg/vect/vect-bswap16.c: Add -msse4 on x86, run on all targets,
> expect vectorized 1 loops message on both vect_bswap and sse4_runtime
> targets.
> * gcc.dg/vect/vect-bswap16a.c: New test.
>
> --- gcc/tree-vect-patterns.c.jj 2019-09-20 12:25:48.186387075 +0200
> +++ gcc/tree-vect-patterns.c 2019-10-01 11:29:18.229215895 +0200
> @@ -46,6 +46,8 @@ along with GCC; see the file COPYING3.
> #include "cgraph.h"
> #include "omp-simd-clone.h"
> #include "predict.h"
> +#include "tree-vector-builder.h"
> +#include "vec-perm-indices.h"
>
> /* Return true if we have a useful VR_RANGE range for VAR, storing it
> in *MIN_VALUE and *MAX_VALUE if so. Note the range in the dump files. */
> @@ -2168,24 +2170,107 @@ vect_recog_rotate_pattern (stmt_vec_info
> enum vect_def_type dt;
> optab optab1, optab2;
> edge ext_def = NULL;
> + bool bswap16_p = false;
>
> - if (!is_gimple_assign (last_stmt))
> - return NULL;
> -
> - rhs_code = gimple_assign_rhs_code (last_stmt);
> - switch (rhs_code)
> + if (is_gimple_assign (last_stmt))
> {
> - case LROTATE_EXPR:
> - case RROTATE_EXPR:
> - break;
> - default:
> - return NULL;
> + rhs_code = gimple_assign_rhs_code (last_stmt);
> + switch (rhs_code)
> + {
> + case LROTATE_EXPR:
> + case RROTATE_EXPR:
> + break;
> + default:
> + return NULL;
> + }
> +
> + lhs = gimple_assign_lhs (last_stmt);
> + oprnd0 = gimple_assign_rhs1 (last_stmt);
> + type = TREE_TYPE (oprnd0);
> + oprnd1 = gimple_assign_rhs2 (last_stmt);
> + }
> + else if (gimple_call_builtin_p (last_stmt, BUILT_IN_BSWAP16))
> + {
> + /* __builtin_bswap16 (x) is another form of x r>> 8.
> + The vectorizer has bswap support, but only if the argument isn't
> + promoted. */
> + lhs = gimple_call_lhs (last_stmt);
> + oprnd0 = gimple_call_arg (last_stmt, 0);
> + type = TREE_TYPE (oprnd0);
> + if (TYPE_PRECISION (TREE_TYPE (lhs)) != 16
> + || TYPE_PRECISION (type) <= 16
> + || TREE_CODE (oprnd0) != SSA_NAME
> + || BITS_PER_UNIT != 8
> + || !TYPE_UNSIGNED (TREE_TYPE (lhs)))
> + return NULL;
> +
> + stmt_vec_info def_stmt_info;
> + if (!vect_is_simple_use (oprnd0, vinfo, &dt, &def_stmt_info,
> &def_stmt))
> + return NULL;
> +
> + if (dt != vect_internal_def)
> + return NULL;
> +
> + if (gimple_assign_cast_p (def_stmt))
> + {
> + def = gimple_assign_rhs1 (def_stmt);
> + if (INTEGRAL_TYPE_P (TREE_TYPE (def))
> + && TYPE_PRECISION (TREE_TYPE (def)) == 16)
> + oprnd0 = def;
> + }
> +
> + type = TREE_TYPE (lhs);
> + vectype = get_vectype_for_scalar_type (type);
> + if (vectype == NULL_TREE)
> + return NULL;
> +
> + if (tree char_vectype = get_same_sized_vectype (char_type_node,
> vectype))
> + {
> + /* The encoding uses one stepped pattern for each byte in the
> + 16-bit word. */
> + vec_perm_builder elts (TYPE_VECTOR_SUBPARTS (char_vectype), 2, 3);
> + for (unsigned i = 0; i < 3; ++i)
> + for (unsigned j = 0; j < 2; ++j)
> + elts.quick_push ((i + 1) * 2 - j - 1);
> +
> + vec_perm_indices indices (elts, 1,
> + TYPE_VECTOR_SUBPARTS (char_vectype));
> + if (can_vec_perm_const_p (TYPE_MODE (char_vectype), indices))
> + {
> + /* vectorizable_bswap can handle the __builtin_bswap16 if we
> + undo the argument promotion. */
> + if (!useless_type_conversion_p (type, TREE_TYPE (oprnd0)))
> + {
> + def = vect_recog_temp_ssa_var (type, NULL);
> + def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd0);
> + append_pattern_def_seq (stmt_vinfo, def_stmt);
> + oprnd0 = def;
> + }
> +
> + /* Pattern detected. */
> + vect_pattern_detected ("vect_recog_rotate_pattern", last_stmt);
> +
> + *type_out = vectype;
> +
> + /* Pattern supported. Create a stmt to be used to replace the
> + pattern, with the unpromoted argument. */
> + var = vect_recog_temp_ssa_var (type, NULL);
> + pattern_stmt = gimple_build_call (gimple_call_fndecl (last_stmt),
> + 1, oprnd0);
> + gimple_call_set_lhs (pattern_stmt, var);
> + gimple_call_set_fntype (as_a <gcall *> (pattern_stmt),
> + gimple_call_fntype (last_stmt));
> + return pattern_stmt;
> + }
> + }
> +
> + oprnd1 = build_int_cst (integer_type_node, 8);
> + rhs_code = LROTATE_EXPR;
> + bswap16_p = true;
> }
> + else
> + return NULL;
>
> - lhs = gimple_assign_lhs (last_stmt);
> - oprnd0 = gimple_assign_rhs1 (last_stmt);
> - type = TREE_TYPE (oprnd0);
> - oprnd1 = gimple_assign_rhs2 (last_stmt);
> if (TREE_CODE (oprnd0) != SSA_NAME
> || TYPE_PRECISION (TREE_TYPE (lhs)) != TYPE_PRECISION (type)
> || !INTEGRAL_TYPE_P (type)
> @@ -2210,14 +2295,39 @@ vect_recog_rotate_pattern (stmt_vec_info
> optab1 = optab_for_tree_code (rhs_code, vectype, optab_vector);
> if (optab1
> && optab_handler (optab1, TYPE_MODE (vectype)) != CODE_FOR_nothing)
> - return NULL;
> + {
> + use_rotate:
> + if (bswap16_p)
> + {
> + if (!useless_type_conversion_p (type, TREE_TYPE (oprnd0)))
> + {
> + def = vect_recog_temp_ssa_var (type, NULL);
> + def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd0);
> + append_pattern_def_seq (stmt_vinfo, def_stmt);
> + oprnd0 = def;
> + }
> +
> + /* Pattern detected. */
> + vect_pattern_detected ("vect_recog_rotate_pattern", last_stmt);
> +
> + *type_out = vectype;
> +
> + /* Pattern supported. Create a stmt to be used to replace the
> + pattern. */
> + var = vect_recog_temp_ssa_var (type, NULL);
> + pattern_stmt = gimple_build_assign (var, LROTATE_EXPR, oprnd0,
> + oprnd1);
> + return pattern_stmt;
> + }
> + return NULL;
> + }
>
> if (is_a <bb_vec_info> (vinfo) || dt != vect_internal_def)
> {
> optab2 = optab_for_tree_code (rhs_code, vectype, optab_scalar);
> if (optab2
> && optab_handler (optab2, TYPE_MODE (vectype)) != CODE_FOR_nothing)
> - return NULL;
> + goto use_rotate;
> }
>
> /* If vector/vector or vector/scalar shifts aren't supported by the target,
> @@ -2242,6 +2352,14 @@ vect_recog_rotate_pattern (stmt_vec_info
>
> *type_out = vectype;
>
> + if (bswap16_p && !useless_type_conversion_p (type, TREE_TYPE (oprnd0)))
> + {
> + def = vect_recog_temp_ssa_var (type, NULL);
> + def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd0);
> + append_pattern_def_seq (stmt_vinfo, def_stmt);
> + oprnd0 = def;
> + }
> +
> if (dt == vect_external_def
> && TREE_CODE (oprnd1) == SSA_NAME)
> ext_def = vect_get_external_def_edge (vinfo, oprnd1);
> --- gcc/testsuite/gcc.dg/vect/vect-bswap16.c.jj 2017-11-09
> 20:24:05.957112733 +0100
> +++ gcc/testsuite/gcc.dg/vect/vect-bswap16.c 2019-10-01 12:39:56.993268063
> +0200
> @@ -1,4 +1,4 @@
> -/* { dg-require-effective-target vect_bswap } */
> +/* { dg-additional-options "-msse4" { target sse4_runtime } } */
>
> #include "tree-vect.h"
>
> @@ -39,4 +39,4 @@ main (void)
> return 0;
> }
>
> -/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
> +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target
> { vect_bswap || sse4_runtime } } } } */
> --- gcc/testsuite/gcc.dg/vect/vect-bswap16a.c.jj 2019-10-01
> 12:35:29.965294527 +0200
> +++ gcc/testsuite/gcc.dg/vect/vect-bswap16a.c 2019-10-01 12:40:45.700533603
> +0200
> @@ -0,0 +1,5 @@
> +/* { dg-additional-options "-msse2 -mno-sse3" { target sse2_runtime } } */
> +
> +#include "vect-bswap16.c"
> +
> +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target
> { vect_shift } } } } */
>
>
> Jakub
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg,
Germany; GF: Felix Imendörffer; HRB 247165 (AG München)
