public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: Richard Guenther <richard.guenther@gmail.com>
To: Jakub Jelinek <jakub@redhat.com>
Cc: Richard Guenther <rguenther@suse.de>,
	Ira Rosen <ira.rosen@linaro.org>,
		Uros Bizjak <ubizjak@gmail.com>,
	Richard Henderson <rth@redhat.com>,
	gcc-patches@gcc.gnu.org
Subject: Re: [PATCH] Gather vectorization (PR tree-optimization/50789, take 2)
Date: Fri, 04 Nov 2011 23:54:00 -0000	[thread overview]
Message-ID: <CAFiYyc1Gg85Rb5RJdZkO+e-P+4VC1C3YOSetsPBF-nSaXyaYbg@mail.gmail.com> (raw)
In-Reply-To: <20111104132231.GN1052@tyan-ft48-01.lab.bos.redhat.com>

On Fri, Nov 4, 2011 at 2:22 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> On Fri, Nov 04, 2011 at 12:21:49PM +0100, Richard Guenther wrote:
>> Ok.  I guess it's ok to use builtins for now - I didn't think of
>> the memory reference issue ;)
>
> Based on IRC discussion I'm posting an updated patch (both former
> patches (base and incremental) in one).
>
> I'm now using expr_invariant_in_loop_p instead of chrec_contains*,
> which nicely handles also the is_gimple_min_invariant case,
> and I've added several comments and fixed the MEM_REF offset
> folding.  Smoke tested on the *gather* testcases, will do full
> bootstrap/regtest soon.

Ok for the vectorizer pieces, I'll defer to x86 maintainers for the target bits.

Thanks,
Richard.

> 2011-11-04  Jakub Jelinek  <jakub@redhat.com>
>
>        PR tree-optimization/50789
>        * tree-vect-stmts.c (process_use): Add force argument, avoid
>        exist_non_indexing_operands_for_use_p check if true.
>        (vect_mark_stmts_to_be_vectorized): Adjust callers.  Handle
>        STMT_VINFO_GATHER_P.
>        (gen_perm_mask): New function.
>        (perm_mask_for_reverse): Use it.
>        (reverse_vec_element): Rename to...
>        (permute_vec_elements): ... this.  Add Y and MASK_VEC arguments,
>        generalize for any permutations.
>        (vectorizable_load): Adjust caller.  Handle STMT_VINFO_GATHER_P.
>        * target.def (TARGET_VECTORIZE_BUILTIN_GATHER): New hook.
>        * doc/tm.texi.in (TARGET_VECTORIZE_BUILTIN_GATHER): Document it.
>        * doc/tm.texi: Regenerate.
>        * tree-data-ref.c (initialize_data_dependence_relation,
>        compute_self_dependence): No longer static.
>        * tree-data-ref.h (initialize_data_dependence_relation,
>        compute_self_dependence): New prototypes.
>        * tree-vect-data-refs.c (vect_check_gather): New function.
>        (vect_analyze_data_refs): Detect possible gather load data
>        refs.
>        * tree-vectorizer.h (struct _stmt_vec_info): Add gather_p field.
>        (STMT_VINFO_GATHER_P): Define.
>        (vect_check_gather): New prototype.
>        * config/i386/i386-builtin-types.def: Add types for alternate
>        gather builtins.
>        * config/i386/sse.md (AVXMODE48P_DI): Remove.
>        (VEC_GATHER_MODE): Rename mode_attr to...
>        (VEC_GATHER_IDXSI): ... this.
>        (VEC_GATHER_IDXDI, VEC_GATHER_SRCDI): New mode_attrs.
>        (avx2_gathersi<mode>, *avx2_gathersi<mode>): Use <VEC_GATHER_IDXSI>
>        instead of <VEC_GATHER_MODE>.
>        (avx2_gatherdi<mode>): Use <VEC_GATHER_IDXDI> instead of
>        <<AVXMODE48P_DI> and <VEC_GATHER_SRCDI> instead of VEC_GATHER_MODE
>        on src and mask operands.
>        (*avx2_gatherdi<mode>): Likewise.  Use VEC_GATHER_MODE iterator
>        instead of AVXMODE48P_DI.
>        (avx2_gatherdi<mode>256, *avx2_gatherdi<mode>256): Removed.
>        * config/i386/i386.c (enum ix86_builtins): Add
>        IX86_BUILTIN_GATHERALTSIV4DF, IX86_BUILTIN_GATHERALTDIV8SF,
>        IX86_BUILTIN_GATHERALTSIV4DI and IX86_BUILTIN_GATHERALTDIV8SI.
>        (ix86_init_mmx_sse_builtins): Create those builtins.
>        (ix86_expand_builtin): Handle those builtins and adjust expansions
>        of other gather builtins.
>        (ix86_vectorize_builtin_gather): New function.
>        (TARGET_VECTORIZE_BUILTIN_GATHER): Define.
>
>        * gcc.target/i386/avx2-gather-1.c: New test.
>        * gcc.target/i386/avx2-gather-2.c: New test.
>        * gcc.target/i386/avx2-gather-3.c: New test.
>        * gcc.target/i386/avx2-gather-4.c: New test.
>
> --- gcc/tree-vect-stmts.c.jj    2011-11-04 08:52:19.000000000 +0100
> +++ gcc/tree-vect-stmts.c       2011-11-04 08:54:11.000000000 +0100
> @@ -332,6 +332,8 @@ exist_non_indexing_operands_for_use_p (t
>    - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt
>      that defined USE.  This is done by calling mark_relevant and passing it
>      the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant).
> +   - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't
> +     be performed.
>
>    Outputs:
>    Generally, LIVE_P and RELEVANT are used to define the liveness and
> @@ -351,7 +353,8 @@ exist_non_indexing_operands_for_use_p (t
>
>  static bool
>  process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p,
> -            enum vect_relevant relevant, VEC(gimple,heap) **worklist)
> +            enum vect_relevant relevant, VEC(gimple,heap) **worklist,
> +            bool force)
>  {
>   struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
>   stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
> @@ -363,7 +366,7 @@ process_use (gimple stmt, tree use, loop
>
>   /* case 1: we are only interested in uses that need to be vectorized.  Uses
>      that are used for address computation are not considered relevant.  */
> -  if (!exist_non_indexing_operands_for_use_p (use, stmt))
> +  if (!force && !exist_non_indexing_operands_for_use_p (use, stmt))
>      return true;
>
>   if (!vect_is_simple_use (use, loop_vinfo, NULL, &def_stmt, &def, &dt))
> @@ -646,7 +649,7 @@ vect_mark_stmts_to_be_vectorized (loop_v
>             break;
>         }
>
> -      if (is_pattern_stmt_p (vinfo_for_stmt (stmt)))
> +      if (is_pattern_stmt_p (stmt_vinfo))
>         {
>           /* Pattern statements are not inserted into the code, so
>              FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we
> @@ -660,9 +663,9 @@ vect_mark_stmts_to_be_vectorized (loop_v
>              if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op))
>                {
>                  if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo,
> -                                   live_p, relevant, &worklist)
> +                                   live_p, relevant, &worklist, false)
>                      || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo,
> -                                      live_p, relevant, &worklist))
> +                                      live_p, relevant, &worklist, false))
>                    {
>                      VEC_free (gimple, heap, worklist);
>                      return false;
> @@ -673,7 +676,7 @@ vect_mark_stmts_to_be_vectorized (loop_v
>                 {
>                  op = gimple_op (stmt, i);
>                   if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
> -                                    &worklist))
> +                                   &worklist, false))
>                     {
>                       VEC_free (gimple, heap, worklist);
>                       return false;
> @@ -686,7 +689,7 @@ vect_mark_stmts_to_be_vectorized (loop_v
>                 {
>                   tree arg = gimple_call_arg (stmt, i);
>                   if (!process_use (stmt, arg, loop_vinfo, live_p, relevant,
> -                                    &worklist))
> +                                   &worklist, false))
>                     {
>                       VEC_free (gimple, heap, worklist);
>                       return false;
> @@ -699,12 +702,25 @@ vect_mark_stmts_to_be_vectorized (loop_v
>           {
>             tree op = USE_FROM_PTR (use_p);
>             if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
> -                              &worklist))
> +                             &worklist, false))
>               {
>                 VEC_free (gimple, heap, worklist);
>                 return false;
>               }
>           }
> +
> +      if (STMT_VINFO_GATHER_P (stmt_vinfo))
> +       {
> +         tree off;
> +         tree decl = vect_check_gather (stmt, loop_vinfo, NULL, &off, NULL);
> +         gcc_assert (decl);
> +         if (!process_use (stmt, off, loop_vinfo, live_p, relevant,
> +                           &worklist, true))
> +           {
> +             VEC_free (gimple, heap, worklist);
> +             return false;
> +           }
> +       }
>     } /* while worklist */
>
>   VEC_free (gimple, heap, worklist);
> @@ -4142,23 +4158,17 @@ vectorizable_store (gimple stmt, gimple_
>   return true;
>  }
>
> -/* Given a vector type VECTYPE returns a builtin DECL to be used
> -   for vector permutation and returns the mask that implements
> -   reversal of the vector elements.  If that is impossible to do,
> -   returns NULL.  */
> +/* Given a vector type VECTYPE and permutation SEL returns
> +   the VECTOR_CST mask that implements the permutation of the
> +   vector elements.  If that is impossible to do, returns NULL.  */
>
>  static tree
> -perm_mask_for_reverse (tree vectype)
> +gen_perm_mask (tree vectype, unsigned char *sel)
>  {
>   tree mask_elt_type, mask_type, mask_vec;
>   int i, nunits;
> -  unsigned char *sel;
>
>   nunits = TYPE_VECTOR_SUBPARTS (vectype);
> -  sel = XALLOCAVEC (unsigned char, nunits);
> -
> -  for (i = 0; i < nunits; ++i)
> -    sel[i] = nunits - 1 - i;
>
>   if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel))
>     return NULL;
> @@ -4169,33 +4179,52 @@ perm_mask_for_reverse (tree vectype)
>   mask_type = get_vectype_for_scalar_type (mask_elt_type);
>
>   mask_vec = NULL;
> -  for (i = 0; i < nunits; i++)
> -    mask_vec = tree_cons (NULL, build_int_cst (mask_elt_type, i), mask_vec);
> +  for (i = nunits - 1; i >= 0; i--)
> +    mask_vec = tree_cons (NULL, build_int_cst (mask_elt_type, sel[i]),
> +                         mask_vec);
>   mask_vec = build_vector (mask_type, mask_vec);
>
>   return mask_vec;
>  }
>
> -/* Given a vector variable X, that was generated for the scalar LHS of
> -   STMT, generate instructions to reverse the vector elements of X,
> -   insert them a *GSI and return the permuted vector variable.  */
> +/* Given a vector type VECTYPE returns the VECTOR_CST mask that implements
> +   reversal of the vector elements.  If that is impossible to do,
> +   returns NULL.  */
>
>  static tree
> -reverse_vec_elements (tree x, gimple stmt, gimple_stmt_iterator *gsi)
> +perm_mask_for_reverse (tree vectype)
> +{
> +  int i, nunits;
> +  unsigned char *sel;
> +
> +  nunits = TYPE_VECTOR_SUBPARTS (vectype);
> +  sel = XALLOCAVEC (unsigned char, nunits);
> +
> +  for (i = 0; i < nunits; ++i)
> +    sel[i] = nunits - 1 - i;
> +
> +  return gen_perm_mask (vectype, sel);
> +}
> +
> +/* Given a vector variable X and Y, that was generated for the scalar
> +   STMT, generate instructions to permute the vector elements of X and Y
> +   using permutation mask MASK_VEC, insert them at *GSI and return the
> +   permuted vector variable.  */
> +
> +static tree
> +permute_vec_elements (tree x, tree y, tree mask_vec, gimple stmt,
> +                     gimple_stmt_iterator *gsi)
>  {
>   tree vectype = TREE_TYPE (x);
> -  tree mask_vec, perm_dest, data_ref;
> +  tree perm_dest, data_ref;
>   gimple perm_stmt;
>
> -  mask_vec = perm_mask_for_reverse (vectype);
> -
>   perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype);
> +  data_ref = make_ssa_name (perm_dest, NULL);
>
>   /* Generate the permute statement.  */
> -  perm_stmt = gimple_build_assign_with_ops3 (VEC_PERM_EXPR, perm_dest,
> -                                            x, x, mask_vec);
> -  data_ref = make_ssa_name (perm_dest, perm_stmt);
> -  gimple_set_lhs (perm_stmt, data_ref);
> +  perm_stmt = gimple_build_assign_with_ops3 (VEC_PERM_EXPR, data_ref,
> +                                            x, y, mask_vec);
>   vect_finish_stmt_generation (stmt, perm_stmt, gsi);
>
>   return data_ref;
> @@ -4254,6 +4283,10 @@ vectorizable_load (gimple stmt, gimple_s
>   bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
>   int vf;
>   tree aggr_type;
> +  tree gather_base = NULL_TREE, gather_off = NULL_TREE;
> +  tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE;
> +  int gather_scale = 1;
> +  enum vect_def_type gather_dt = vect_unknown_def_type;
>
>   if (loop_vinfo)
>     {
> @@ -4334,7 +4367,7 @@ vectorizable_load (gimple stmt, gimple_s
>     {
>       strided_load = true;
>       /* FORNOW */
> -      gcc_assert (! nested_in_vect_loop);
> +      gcc_assert (! nested_in_vect_loop && !STMT_VINFO_GATHER_P (stmt_info));
>
>       first_stmt = GROUP_FIRST_ELEMENT (stmt_info);
>       if (!slp && !PURE_SLP_STMT (stmt_info))
> @@ -4349,7 +4382,7 @@ vectorizable_load (gimple stmt, gimple_s
>
>   if (negative)
>     {
> -      gcc_assert (!strided_load);
> +      gcc_assert (!strided_load && !STMT_VINFO_GATHER_P (stmt_info));
>       alignment_support_scheme = vect_supportable_dr_alignment (dr, false);
>       if (alignment_support_scheme != dr_aligned
>          && alignment_support_scheme != dr_unaligned_supported)
> @@ -4366,6 +4399,23 @@ vectorizable_load (gimple stmt, gimple_s
>        }
>     }
>
> +  if (STMT_VINFO_GATHER_P (stmt_info))
> +    {
> +      gimple def_stmt;
> +      tree def;
> +      gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base,
> +                                      &gather_off, &gather_scale);
> +      gcc_assert (gather_decl);
> +      if (!vect_is_simple_use_1 (gather_off, loop_vinfo, bb_vinfo,
> +                                &def_stmt, &def, &gather_dt,
> +                                &gather_off_vectype))
> +       {
> +         if (vect_print_dump_info (REPORT_DETAILS))
> +           fprintf (vect_dump, "gather index use not simple.");
> +         return false;
> +       }
> +    }
> +
>   if (!vec_stmt) /* transformation not required.  */
>     {
>       STMT_VINFO_TYPE (stmt_info) = load_vec_info_type;
> @@ -4378,6 +4428,161 @@ vectorizable_load (gimple stmt, gimple_s
>
>   /** Transform.  **/
>
> +  if (STMT_VINFO_GATHER_P (stmt_info))
> +    {
> +      tree vec_oprnd0 = NULL_TREE, op;
> +      tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl));
> +      tree rettype, srctype, ptrtype, idxtype, masktype, scaletype;
> +      tree ptr, mask, var, scale, perm_mask = NULL_TREE, prev_res = NULL_TREE;
> +      edge pe = loop_preheader_edge (loop);
> +      gimple_seq seq;
> +      basic_block new_bb;
> +      enum { NARROW, NONE, WIDEN } modifier;
> +      int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype);
> +
> +      if (nunits == gather_off_nunits)
> +       modifier = NONE;
> +      else if (nunits == gather_off_nunits / 2)
> +       {
> +         unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits);
> +         modifier = WIDEN;
> +
> +         for (i = 0; i < gather_off_nunits; ++i)
> +           sel[i] = i | nunits;
> +
> +         perm_mask = gen_perm_mask (gather_off_vectype, sel);
> +         gcc_assert (perm_mask != NULL_TREE);
> +       }
> +      else if (nunits == gather_off_nunits * 2)
> +       {
> +         unsigned char *sel = XALLOCAVEC (unsigned char, nunits);
> +         modifier = NARROW;
> +
> +         for (i = 0; i < nunits; ++i)
> +           sel[i] = i < gather_off_nunits
> +                    ? i : i + nunits - gather_off_nunits;
> +
> +         perm_mask = gen_perm_mask (vectype, sel);
> +         gcc_assert (perm_mask != NULL_TREE);
> +         ncopies *= 2;
> +       }
> +      else
> +       gcc_unreachable ();
> +
> +      rettype = TREE_TYPE (TREE_TYPE (gather_decl));
> +      srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
> +      ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
> +      idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
> +      masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
> +      scaletype = TREE_VALUE (arglist);
> +      gcc_checking_assert (types_compatible_p (srctype, rettype)
> +                          && types_compatible_p (srctype, masktype));
> +
> +      vec_dest = vect_create_destination_var (scalar_dest, vectype);
> +
> +      ptr = fold_convert (ptrtype, gather_base);
> +      if (!is_gimple_min_invariant (ptr))
> +       {
> +         ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE);
> +         new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
> +         gcc_assert (!new_bb);
> +       }
> +
> +      /* Currently we support only unconditional gather loads,
> +        so mask should be all ones.  */
> +      if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE)
> +       mask = build_int_cst (TREE_TYPE (masktype), -1);
> +      else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype)))
> +       {
> +         REAL_VALUE_TYPE r;
> +         long tmp[6];
> +         for (j = 0; j < 6; ++j)
> +           tmp[j] = -1;
> +         real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype)));
> +         mask = build_real (TREE_TYPE (masktype), r);
> +       }
> +      else
> +       gcc_unreachable ();
> +      mask = build_vector_from_val (masktype, mask);
> +      mask = vect_init_vector (stmt, mask, masktype, NULL);
> +
> +      scale = build_int_cst (scaletype, gather_scale);
> +
> +      prev_stmt_info = NULL;
> +      for (j = 0; j < ncopies; ++j)
> +       {
> +         if (modifier == WIDEN && (j & 1))
> +           op = permute_vec_elements (vec_oprnd0, vec_oprnd0,
> +                                      perm_mask, stmt, gsi);
> +         else if (j == 0)
> +           op = vec_oprnd0
> +             = vect_get_vec_def_for_operand (gather_off, stmt, NULL);
> +         else
> +           op = vec_oprnd0
> +             = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0);
> +
> +         if (!useless_type_conversion_p (idxtype, TREE_TYPE (op)))
> +           {
> +             gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op))
> +                         == TYPE_VECTOR_SUBPARTS (idxtype));
> +             var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL);
> +             add_referenced_var (var);
> +             var = make_ssa_name (var, NULL);
> +             op = build1 (VIEW_CONVERT_EXPR, idxtype, op);
> +             new_stmt
> +               = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var,
> +                                               op, NULL_TREE);
> +             vect_finish_stmt_generation (stmt, new_stmt, gsi);
> +             op = var;
> +           }
> +
> +         new_stmt
> +           = gimple_build_call (gather_decl, 5, mask, ptr, op, mask, scale);
> +
> +         if (!useless_type_conversion_p (vectype, rettype))
> +           {
> +             gcc_assert (TYPE_VECTOR_SUBPARTS (vectype)
> +                         == TYPE_VECTOR_SUBPARTS (rettype));
> +             var = vect_get_new_vect_var (rettype, vect_simple_var, NULL);
> +             add_referenced_var (var);
> +             op = make_ssa_name (var, new_stmt);
> +             gimple_call_set_lhs (new_stmt, op);
> +             vect_finish_stmt_generation (stmt, new_stmt, gsi);
> +             var = make_ssa_name (vec_dest, NULL);
> +             op = build1 (VIEW_CONVERT_EXPR, vectype, op);
> +             new_stmt
> +               = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, op,
> +                                               NULL_TREE);
> +           }
> +         else
> +           {
> +             var = make_ssa_name (vec_dest, new_stmt);
> +             gimple_call_set_lhs (new_stmt, var);
> +           }
> +
> +         vect_finish_stmt_generation (stmt, new_stmt, gsi);
> +
> +         if (modifier == NARROW)
> +           {
> +             if ((j & 1) == 0)
> +               {
> +                 prev_res = var;
> +                 continue;
> +               }
> +             var = permute_vec_elements (prev_res, var,
> +                                         perm_mask, stmt, gsi);
> +             new_stmt = SSA_NAME_DEF_STMT (var);
> +           }
> +
> +         if (prev_stmt_info == NULL)
> +           STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
> +         else
> +           STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
> +         prev_stmt_info = vinfo_for_stmt (new_stmt);
> +       }
> +      return true;
> +    }
> +
>   if (strided_load)
>     {
>       first_stmt = GROUP_FIRST_ELEMENT (stmt_info);
> @@ -4769,7 +4974,9 @@ vectorizable_load (gimple stmt, gimple_s
>
>              if (negative)
>                {
> -                 new_temp = reverse_vec_elements (new_temp, stmt, gsi);
> +                 tree perm_mask = perm_mask_for_reverse (vectype);
> +                 new_temp = permute_vec_elements (new_temp, new_temp,
> +                                                  perm_mask, stmt, gsi);
>                  new_stmt = SSA_NAME_DEF_STMT (new_temp);
>                }
>
> --- gcc/target.def.jj   2011-11-04 08:52:19.000000000 +0100
> +++ gcc/target.def      2011-11-04 08:53:13.000000000 +0100
> @@ -1021,6 +1021,14 @@ DEFHOOK
>  (void),
>  default_autovectorize_vector_sizes)
>
> +/* Target builtin that implements vector gather operation.  */
> +DEFHOOK
> +(builtin_gather,
> + "",
> + tree,
> + (const_tree mem_vectype, const_tree index_type, int scale),
> + NULL)
> +
>  HOOK_VECTOR_END (vectorize)
>
>  #undef HOOK_PREFIX
> --- gcc/tree-data-ref.c.jj      2011-11-04 08:52:19.000000000 +0100
> +++ gcc/tree-data-ref.c 2011-11-04 08:53:13.000000000 +0100
> @@ -1351,13 +1351,11 @@ dr_may_alias_p (const struct data_refere
>   return refs_may_alias_p (addr_a, addr_b);
>  }
>
> -static void compute_self_dependence (struct data_dependence_relation *);
> -
>  /* Initialize a data dependence relation between data accesses A and
>    B.  NB_LOOPS is the number of loops surrounding the references: the
>    size of the classic distance/direction vectors.  */
>
> -static struct data_dependence_relation *
> +struct data_dependence_relation *
>  initialize_data_dependence_relation (struct data_reference *a,
>                                     struct data_reference *b,
>                                     VEC (loop_p, heap) *loop_nest)
> @@ -4121,7 +4119,7 @@ compute_affine_dependence (struct data_d
>  /* This computes the dependence relation for the same data
>    reference into DDR.  */
>
> -static void
> +void
>  compute_self_dependence (struct data_dependence_relation *ddr)
>  {
>   unsigned int i;
> --- gcc/tree-data-ref.h.jj      2011-11-04 08:52:19.000000000 +0100
> +++ gcc/tree-data-ref.h 2011-11-04 13:22:28.000000000 +0100
> @@ -1,5 +1,5 @@
>  /* Data references and dependences detectors.
> -   Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
> +   Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011
>    Free Software Foundation, Inc.
>    Contributed by Sebastian Pop <pop@cri.ensmp.fr>
>
> @@ -423,6 +423,9 @@ extern bool graphite_find_data_reference
>                                                   VEC (data_reference_p, heap) **);
>  struct data_reference *create_data_ref (loop_p, loop_p, tree, gimple, bool);
>  extern bool find_loop_nest (struct loop *, VEC (loop_p, heap) **);
> +extern struct data_dependence_relation *initialize_data_dependence_relation
> +     (struct data_reference *, struct data_reference *, VEC (loop_p, heap) *);
> +extern void compute_self_dependence (struct data_dependence_relation *);
>  extern void compute_all_dependences (VEC (data_reference_p, heap) *,
>                                     VEC (ddr_p, heap) **, VEC (loop_p, heap) *,
>                                     bool);
> --- gcc/doc/tm.texi.in.jj       2011-11-04 08:52:19.000000000 +0100
> +++ gcc/doc/tm.texi.in  2011-11-04 08:53:13.000000000 +0100
> @@ -5696,6 +5696,14 @@ mode returned by @code{TARGET_VECTORIZE_
>  The default is zero which means to not iterate over other vector sizes.
>  @end deftypefn
>
> +@hook TARGET_VECTORIZE_BUILTIN_GATHER
> +Target builtin that implements vector gather operation.  @var{mem_vectype}
> +is the vector type of the load and @var{index_type} is scalar type of
> +the index, scaled by @var{scale}.
> +The default is @code{NULL_TREE} which means to not vectorize gather
> +loads.
> +@end deftypefn
> +
>  @node Anchored Addresses
>  @section Anchored Addresses
>  @cindex anchored addresses
> --- gcc/doc/tm.texi.jj  2011-11-04 08:52:19.000000000 +0100
> +++ gcc/doc/tm.texi     2011-11-04 08:53:13.000000000 +0100
> @@ -5758,6 +5758,14 @@ mode returned by @code{TARGET_VECTORIZE_
>  The default is zero which means to not iterate over other vector sizes.
>  @end deftypefn
>
> +@deftypefn {Target Hook} tree TARGET_VECTORIZE_BUILTIN_GATHER (const_tree @var{mem_vectype}, const_tree @var{index_type}, int @var{scale})
> +Target builtin that implements vector gather operation.  @var{mem_vectype}
> +is the vector type of the load and @var{index_type} is scalar type of
> +the index, scaled by @var{scale}.
> +The default is @code{NULL_TREE} which means to not vectorize gather
> +loads.
> +@end deftypefn
> +
>  @node Anchored Addresses
>  @section Anchored Addresses
>  @cindex anchored addresses
> --- gcc/tree-vect-data-refs.c.jj        2011-11-04 08:52:57.000000000 +0100
> +++ gcc/tree-vect-data-refs.c   2011-11-04 14:10:17.000000000 +0100
> @@ -2497,6 +2497,199 @@ vect_prune_runtime_alias_test_list (loop
>   return true;
>  }
>
> +/* Check whether a non-affine read in stmt is suitable for gather load
> +   and if so, return a builtin decl for that operation.  */
> +
> +tree
> +vect_check_gather (gimple stmt, loop_vec_info loop_vinfo, tree *basep,
> +                  tree *offp, int *scalep)
> +{
> +  HOST_WIDE_INT scale = 1, pbitpos, pbitsize;
> +  struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
> +  stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
> +  struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
> +  tree offtype = NULL_TREE;
> +  tree decl, base, off;
> +  enum machine_mode pmode;
> +  int punsignedp, pvolatilep;
> +
> +  /* The gather builtins need address of the form
> +     loop_invariant + vector * {1, 2, 4, 8}
> +     or
> +     loop_invariant + sign_extend (vector) * { 1, 2, 4, 8 }.
> +     Unfortunately DR_BASE_ADDRESS/DR_OFFSET can be a mixture
> +     of loop invariants/SSA_NAMEs defined in the loop, with casts,
> +     multiplications and additions in it.  To get a vector, we need
> +     a single SSA_NAME that will be defined in the loop and will
> +     contain everything that is not loop invariant and that can be
> +     vectorized.  The following code attempts to find such a preexistng
> +     SSA_NAME OFF and put the loop invariants into a tree BASE
> +     that can be gimplified before the loop.  */
> +  base = get_inner_reference (DR_REF (dr), &pbitsize, &pbitpos, &off,
> +                             &pmode, &punsignedp, &pvolatilep, false);
> +  gcc_assert (base != NULL_TREE && (pbitpos % BITS_PER_UNIT) == 0);
> +
> +  if (TREE_CODE (base) == MEM_REF)
> +    {
> +      if (!integer_zerop (TREE_OPERAND (base, 1)))
> +       {
> +         if (off == NULL_TREE)
> +           {
> +             double_int moff = mem_ref_offset (base);
> +             off = double_int_to_tree (sizetype, moff);
> +           }
> +         else
> +           off = size_binop (PLUS_EXPR, off,
> +                             fold_convert (sizetype, TREE_OPERAND (base, 1)));
> +       }
> +      base = TREE_OPERAND (base, 0);
> +    }
> +  else
> +    base = build_fold_addr_expr (base);
> +
> +  if (off == NULL_TREE)
> +    off = size_zero_node;
> +
> +  /* If base is not loop invariant, either off is 0, then we start with just
> +     the constant offset in the loop invariant BASE and continue with base
> +     as OFF, otherwise give up.
> +     We could handle that case by gimplifying the addition of base + off
> +     into some SSA_NAME and use that as off, but for now punt.  */
> +  if (!expr_invariant_in_loop_p (loop, base))
> +    {
> +      if (!integer_zerop (off))
> +       return NULL_TREE;
> +      off = base;
> +      base = size_int (pbitpos / BITS_PER_UNIT);
> +    }
> +  /* Otherwise put base + constant offset into the loop invariant BASE
> +     and continue with OFF.  */
> +  else
> +    {
> +      base = fold_convert (sizetype, base);
> +      base = size_binop (PLUS_EXPR, base, size_int (pbitpos / BITS_PER_UNIT));
> +    }
> +
> +  /* OFF at this point may be either a SSA_NAME or some tree expression
> +     from get_inner_reference.  Try to peel off loop invariants from it
> +     into BASE as long as possible.  */
> +  STRIP_NOPS (off);
> +  while (offtype == NULL_TREE)
> +    {
> +      enum tree_code code;
> +      tree op0, op1, add = NULL_TREE;
> +
> +      if (TREE_CODE (off) == SSA_NAME)
> +       {
> +         gimple def_stmt = SSA_NAME_DEF_STMT (off);
> +
> +         if (expr_invariant_in_loop_p (loop, off))
> +           return NULL_TREE;
> +
> +         if (gimple_code (def_stmt) != GIMPLE_ASSIGN)
> +           break;
> +
> +         op0 = gimple_assign_rhs1 (def_stmt);
> +         code = gimple_assign_rhs_code (def_stmt);
> +         op1 = gimple_assign_rhs2 (def_stmt);
> +       }
> +      else
> +       {
> +         if (get_gimple_rhs_class (TREE_CODE (off)) == GIMPLE_TERNARY_RHS)
> +           return NULL_TREE;
> +         code = TREE_CODE (off);
> +         extract_ops_from_tree (off, &code, &op0, &op1);
> +       }
> +      switch (code)
> +       {
> +       case POINTER_PLUS_EXPR:
> +       case PLUS_EXPR:
> +         if (expr_invariant_in_loop_p (loop, op0))
> +           {
> +             add = op0;
> +             off = op1;
> +           do_add:
> +             add = fold_convert (sizetype, add);
> +             if (scale != 1)
> +               add = size_binop (MULT_EXPR, add, size_int (scale));
> +             base = size_binop (PLUS_EXPR, base, add);
> +             continue;
> +           }
> +         if (expr_invariant_in_loop_p (loop, op1))
> +           {
> +             add = op1;
> +             off = op0;
> +             goto do_add;
> +           }
> +         break;
> +       case MINUS_EXPR:
> +         if (expr_invariant_in_loop_p (loop, op1))
> +           {
> +             add = fold_convert (sizetype, op1);
> +             add = size_binop (MINUS_EXPR, size_zero_node, add);
> +             off = op0;
> +             goto do_add;
> +           }
> +         break;
> +       case MULT_EXPR:
> +         if (scale == 1 && host_integerp (op1, 0))
> +           {
> +             scale = tree_low_cst (op1, 0);
> +             off = op0;
> +             continue;
> +           }
> +         break;
> +       case SSA_NAME:
> +         off = op0;
> +         continue;
> +       CASE_CONVERT:
> +         if (!POINTER_TYPE_P (TREE_TYPE (op0))
> +             && !INTEGRAL_TYPE_P (TREE_TYPE (op0)))
> +           break;
> +         if (TYPE_PRECISION (TREE_TYPE (op0))
> +             == TYPE_PRECISION (TREE_TYPE (off)))
> +           {
> +             off = op0;
> +             continue;
> +           }
> +         if (TYPE_PRECISION (TREE_TYPE (op0))
> +             < TYPE_PRECISION (TREE_TYPE (off)))
> +           {
> +             off = op0;
> +             offtype = TREE_TYPE (off);
> +             STRIP_NOPS (off);
> +             continue;
> +           }
> +         break;
> +       default:
> +         break;
> +       }
> +      break;
> +    }
> +
> +  /* If at the end OFF still isn't a SSA_NAME or isn't
> +     defined in the loop, punt.  */
> +  if (TREE_CODE (off) != SSA_NAME
> +      || expr_invariant_in_loop_p (loop, off))
> +    return NULL_TREE;
> +
> +  if (offtype == NULL_TREE)
> +    offtype = TREE_TYPE (off);
> +
> +  decl = targetm.vectorize.builtin_gather (STMT_VINFO_VECTYPE (stmt_info),
> +                                          offtype, scale);
> +  if (decl == NULL_TREE)
> +    return NULL_TREE;
> +
> +  if (basep)
> +    *basep = base;
> +  if (offp)
> +    *offp = off;
> +  if (scalep)
> +    *scalep = scale;
> +  return decl;
> +}
> +
>
>  /* Function vect_analyze_data_refs.
>
> @@ -2573,6 +2766,7 @@ vect_analyze_data_refs (loop_vec_info lo
>       gimple stmt;
>       stmt_vec_info stmt_info;
>       tree base, offset, init;
> +      bool gather = false;
>       int vf;
>
>       if (!dr || !DR_REF (dr))
> @@ -2594,22 +2788,51 @@ vect_analyze_data_refs (loop_vec_info lo
>
>       /* Check that analysis of the data-ref succeeded.  */
>       if (!DR_BASE_ADDRESS (dr) || !DR_OFFSET (dr) || !DR_INIT (dr)
> -          || !DR_STEP (dr))
> +         || !DR_STEP (dr))
>         {
> -          if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
> -            {
> -              fprintf (vect_dump, "not vectorized: data ref analysis failed ");
> -              print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
> -            }
> +         /* If target supports vector gather loads, see if they can't
> +            be used.  */
> +         if (loop_vinfo
> +             && DR_IS_READ (dr)
> +             && !TREE_THIS_VOLATILE (DR_REF (dr))
> +             && targetm.vectorize.builtin_gather != NULL
> +             && !nested_in_vect_loop_p (loop, stmt))
> +           {
> +             struct data_reference *newdr
> +               = create_data_ref (NULL, loop_containing_stmt (stmt),
> +                                  DR_REF (dr), stmt, true);
> +             gcc_assert (newdr != NULL && DR_REF (newdr));
> +             if (DR_BASE_ADDRESS (newdr)
> +                 && DR_OFFSET (newdr)
> +                 && DR_INIT (newdr)
> +                 && DR_STEP (newdr)
> +                 && integer_zerop (DR_STEP (newdr)))
> +               {
> +                 dr = newdr;
> +                 gather = true;
> +               }
> +             else
> +               free_data_ref (newdr);
> +           }
>
> -          if (bb_vinfo)
> -            {
> -              STMT_VINFO_VECTORIZABLE (stmt_info) = false;
> -              stop_bb_analysis = true;
> -              continue;
> -            }
> +         if (!gather)
> +           {
> +             if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
> +               {
> +                 fprintf (vect_dump, "not vectorized: data ref analysis "
> +                                     "failed ");
> +                 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
> +               }
> +
> +             if (bb_vinfo)
> +               {
> +                 STMT_VINFO_VECTORIZABLE (stmt_info) = false;
> +                 stop_bb_analysis = true;
> +                 continue;
> +               }
>
> -          return false;
> +             return false;
> +           }
>         }
>
>       if (TREE_CODE (DR_BASE_ADDRESS (dr)) == INTEGER_CST)
> @@ -2625,7 +2848,9 @@ vect_analyze_data_refs (loop_vec_info lo
>               continue;
>             }
>
> -           return false;
> +         if (gather)
> +           free_data_ref (dr);
> +         return false;
>         }
>
>       if (TREE_THIS_VOLATILE (DR_REF (dr)))
> @@ -2666,6 +2891,8 @@ vect_analyze_data_refs (loop_vec_info lo
>               continue;
>             }
>
> +         if (gather)
> +           free_data_ref (dr);
>           return false;
>         }
>
> @@ -2791,6 +3018,8 @@ vect_analyze_data_refs (loop_vec_info lo
>               continue;
>             }
>
> +         if (gather)
> +           free_data_ref (dr);
>           return false;
>         }
>
> @@ -2818,8 +3047,13 @@ vect_analyze_data_refs (loop_vec_info lo
>               stop_bb_analysis = true;
>               continue;
>             }
> -          else
> -            return false;
> +
> +         if (gather)
> +           {
> +             STMT_VINFO_DATA_REF (stmt_info) = NULL;
> +             free_data_ref (dr);
> +           }
> +         return false;
>         }
>
>       /* Adjust the minimal vectorization factor according to the
> @@ -2827,6 +3061,86 @@ vect_analyze_data_refs (loop_vec_info lo
>       vf = TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info));
>       if (vf > *min_vf)
>        *min_vf = vf;
> +
> +      if (gather)
> +       {
> +         unsigned int j, k, n;
> +         struct data_reference *olddr
> +           = VEC_index (data_reference_p, datarefs, i);
> +         VEC (ddr_p, heap) *ddrs = LOOP_VINFO_DDRS (loop_vinfo);
> +         struct data_dependence_relation *ddr, *newddr;
> +         bool bad = false;
> +         tree off;
> +         VEC (loop_p, heap) *nest = LOOP_VINFO_LOOP_NEST (loop_vinfo);
> +
> +         if (!vect_check_gather (stmt, loop_vinfo, NULL, &off, NULL)
> +             || get_vectype_for_scalar_type (TREE_TYPE (off)) == NULL_TREE)
> +           {
> +             if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
> +               {
> +                 fprintf (vect_dump,
> +                          "not vectorized: not suitable for gather ");
> +                 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
> +               }
> +             return false;
> +           }
> +
> +         n = VEC_length (data_reference_p, datarefs) - 1;
> +         for (j = 0, k = i - 1; j < i; j++)
> +           {
> +             ddr = VEC_index (ddr_p, ddrs, k);
> +             gcc_assert (DDR_B (ddr) == olddr);
> +             newddr = initialize_data_dependence_relation (DDR_A (ddr), dr,
> +                                                           nest);
> +             VEC_replace (ddr_p, ddrs, k, newddr);
> +             free_dependence_relation (ddr);
> +             if (!bad
> +                 && DR_IS_WRITE (DDR_A (newddr))
> +                 && DDR_ARE_DEPENDENT (newddr) != chrec_known)
> +               bad = true;
> +             k += --n;
> +           }
> +
> +         k++;
> +         n = k + VEC_length (data_reference_p, datarefs) - i - 1;
> +         for (; k < n; k++)
> +           {
> +             ddr = VEC_index (ddr_p, ddrs, k);
> +             gcc_assert (DDR_A (ddr) == olddr);
> +             newddr = initialize_data_dependence_relation (dr, DDR_B (ddr),
> +                                                           nest);
> +             VEC_replace (ddr_p, ddrs, k, newddr);
> +             free_dependence_relation (ddr);
> +             if (!bad
> +                 && DR_IS_WRITE (DDR_B (newddr))
> +                 && DDR_ARE_DEPENDENT (newddr) != chrec_known)
> +               bad = true;
> +           }
> +
> +         k = VEC_length (ddr_p, ddrs)
> +             - VEC_length (data_reference_p, datarefs) + i;
> +         ddr = VEC_index (ddr_p, ddrs, k);
> +         gcc_assert (DDR_A (ddr) == olddr && DDR_B (ddr) == olddr);
> +         newddr = initialize_data_dependence_relation (dr, dr, nest);
> +         compute_self_dependence (newddr);
> +         VEC_replace (ddr_p, ddrs, k, newddr);
> +         free_dependence_relation (ddr);
> +         VEC_replace (data_reference_p, datarefs, i, dr);
> +
> +         if (bad)
> +           {
> +             if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
> +               {
> +                 fprintf (vect_dump,
> +                          "not vectorized: data dependence conflict"
> +                          " prevents gather");
> +                 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
> +               }
> +             return false;
> +           }
> +
> +         STMT_VINFO_GATHER_P (stmt_info) = true;
> +       }
>     }
>
>   return true;
> --- gcc/tree-vectorizer.h.jj    2011-11-04 08:52:19.000000000 +0100
> +++ gcc/tree-vectorizer.h       2011-11-04 08:53:13.000000000 +0100
> @@ -535,6 +535,9 @@ typedef struct _stmt_vec_info {
>   /* Is this statement vectorizable or should it be skipped in (partial)
>      vectorization.  */
>   bool vectorizable;
> +
> +  /* For loads only, true if this is a gather load.  */
> +  bool gather_p;
>  } *stmt_vec_info;
>
>  /* Access Functions.  */
> @@ -548,6 +551,7 @@ typedef struct _stmt_vec_info {
>  #define STMT_VINFO_VEC_STMT(S)             (S)->vectorized_stmt
>  #define STMT_VINFO_VECTORIZABLE(S)         (S)->vectorizable
>  #define STMT_VINFO_DATA_REF(S)             (S)->data_ref_info
> +#define STMT_VINFO_GATHER_P(S)            (S)->gather_p
>
>  #define STMT_VINFO_DR_BASE_ADDRESS(S)      (S)->dr_base_address
>  #define STMT_VINFO_DR_INIT(S)              (S)->dr_init
> @@ -858,6 +862,8 @@ extern bool vect_analyze_data_refs_align
>  extern bool vect_verify_datarefs_alignment (loop_vec_info, bb_vec_info);
>  extern bool vect_analyze_data_ref_accesses (loop_vec_info, bb_vec_info);
>  extern bool vect_prune_runtime_alias_test_list (loop_vec_info);
> +extern tree vect_check_gather (gimple, loop_vec_info, tree *, tree *,
> +                              int *);
>  extern bool vect_analyze_data_refs (loop_vec_info, bb_vec_info, int *);
>  extern tree vect_create_data_ref_ptr (gimple, tree, struct loop *, tree,
>                                      tree *, gimple_stmt_iterator *,
> --- gcc/config/i386/i386-builtin-types.def.jj   2011-11-04 08:52:19.000000000 +0100
> +++ gcc/config/i386/i386-builtin-types.def      2011-11-04 08:53:13.000000000 +0100
> @@ -432,20 +432,24 @@ DEF_FUNCTION_TYPE (V8QI, QI, QI, QI, QI,
>
>  DEF_FUNCTION_TYPE (V2DF, V2DF, PCDOUBLE, V4SI, V2DF, INT)
>  DEF_FUNCTION_TYPE (V4DF, V4DF, PCDOUBLE, V4SI, V4DF, INT)
> +DEF_FUNCTION_TYPE (V4DF, V4DF, PCDOUBLE, V8SI, V4DF, INT)
>  DEF_FUNCTION_TYPE (V2DF, V2DF, PCDOUBLE, V2DI, V2DF, INT)
>  DEF_FUNCTION_TYPE (V4DF, V4DF, PCDOUBLE, V4DI, V4DF, INT)
>  DEF_FUNCTION_TYPE (V4SF, V4SF, PCFLOAT, V4SI, V4SF, INT)
>  DEF_FUNCTION_TYPE (V8SF, V8SF, PCFLOAT, V8SI, V8SF, INT)
>  DEF_FUNCTION_TYPE (V4SF, V4SF, PCFLOAT, V2DI, V4SF, INT)
>  DEF_FUNCTION_TYPE (V4SF, V4SF, PCFLOAT, V4DI, V4SF, INT)
> +DEF_FUNCTION_TYPE (V8SF, V8SF, PCFLOAT, V4DI, V8SF, INT)
>  DEF_FUNCTION_TYPE (V2DI, V2DI, PCINT64, V4SI, V2DI, INT)
>  DEF_FUNCTION_TYPE (V4DI, V4DI, PCINT64, V4SI, V4DI, INT)
> +DEF_FUNCTION_TYPE (V4DI, V4DI, PCINT64, V8SI, V4DI, INT)
>  DEF_FUNCTION_TYPE (V2DI, V2DI, PCINT64, V2DI, V2DI, INT)
>  DEF_FUNCTION_TYPE (V4DI, V4DI, PCINT64, V4DI, V4DI, INT)
>  DEF_FUNCTION_TYPE (V4SI, V4SI, PCINT, V4SI, V4SI, INT)
>  DEF_FUNCTION_TYPE (V8SI, V8SI, PCINT, V8SI, V8SI, INT)
>  DEF_FUNCTION_TYPE (V4SI, V4SI, PCINT, V2DI, V4SI, INT)
>  DEF_FUNCTION_TYPE (V4SI, V4SI, PCINT, V4DI, V4SI, INT)
> +DEF_FUNCTION_TYPE (V8SI, V8SI, PCINT, V4DI, V8SI, INT)
>
>  DEF_FUNCTION_TYPE_ALIAS (V2DF_FTYPE_V2DF, ROUND)
>  DEF_FUNCTION_TYPE_ALIAS (V4DF_FTYPE_V4DF, ROUND)
> --- gcc/config/i386/sse.md.jj   2011-11-04 08:52:19.000000000 +0100
> +++ gcc/config/i386/sse.md      2011-11-04 12:48:16.000000000 +0100
> @@ -316,14 +316,6 @@ (define_mode_attr i128
>  ;; Mix-n-match
>  (define_mode_iterator AVX256MODE2P [V8SI V8SF V4DF])
>
> -(define_mode_iterator AVXMODE48P_DI
> -                     [V2DI V2DF V4DI V4DF V4SF V4SI])
> -(define_mode_attr AVXMODE48P_DI
> -                     [(V2DI "V2DI") (V2DF "V2DI")
> -                      (V4DI "V4DI") (V4DF "V4DI")
> -                      (V4SI "V2DI") (V4SF "V2DI")
> -                      (V8SI "V4DI") (V8SF "V4DI")])
> -
>  (define_mode_iterator FMAMODE [SF DF V4SF V2DF V8SF V4DF])
>
>  ;; Mapping of immediate bits for blend instructions
> @@ -12516,11 +12508,21 @@ (define_insn "vcvtps2ph256"
>  ;; For gather* insn patterns
>  (define_mode_iterator VEC_GATHER_MODE
>                      [V2DI V2DF V4DI V4DF V4SI V4SF V8SI V8SF])
> -(define_mode_attr VEC_GATHER_MODE
> +(define_mode_attr VEC_GATHER_IDXSI
>                      [(V2DI "V4SI") (V2DF "V4SI")
>                       (V4DI "V4SI") (V4DF "V4SI")
>                       (V4SI "V4SI") (V4SF "V4SI")
>                       (V8SI "V8SI") (V8SF "V8SI")])
> +(define_mode_attr VEC_GATHER_IDXDI
> +                     [(V2DI "V2DI") (V2DF "V2DI")
> +                      (V4DI "V4DI") (V4DF "V4DI")
> +                      (V4SI "V2DI") (V4SF "V2DI")
> +                      (V8SI "V4DI") (V8SF "V4DI")])
> +(define_mode_attr VEC_GATHER_SRCDI
> +                     [(V2DI "V2DI") (V2DF "V2DF")
> +                      (V4DI "V4DI") (V4DF "V4DF")
> +                      (V4SI "V4SI") (V4SF "V4SF")
> +                      (V8SI "V4SI") (V8SF "V4SF")])
>
>  (define_expand "avx2_gathersi<mode>"
>   [(parallel [(set (match_operand:VEC_GATHER_MODE 0 "register_operand" "")
> @@ -12529,7 +12531,8 @@ (define_expand "avx2_gathersi<mode>"
>                      (mem:<ssescalarmode>
>                        (match_par_dup 7
>                          [(match_operand 2 "vsib_address_operand" "")
> -                          (match_operand:<VEC_GATHER_MODE> 3 "register_operand" "")
> +                          (match_operand:<VEC_GATHER_IDXSI>
> +                             3 "register_operand" "")
>                           (match_operand:SI 5 "const1248_operand " "")]))
>                      (mem:BLK (scratch))
>                      (match_operand:VEC_GATHER_MODE 4 "register_operand" "")]
> @@ -12549,7 +12552,7 @@ (define_insn "*avx2_gathersi<mode>"
>           (match_operator:<ssescalarmode> 7 "vsib_mem_operator"
>             [(unspec:P
>                [(match_operand:P 3 "vsib_address_operand" "p")
> -                (match_operand:<VEC_GATHER_MODE> 4 "register_operand" "x")
> +                (match_operand:<VEC_GATHER_IDXSI> 4 "register_operand" "x")
>                 (match_operand:SI 6 "const1248_operand" "n")]
>                UNSPEC_VSIBADDR)])
>           (mem:BLK (scratch))
> @@ -12565,14 +12568,16 @@ (define_insn "*avx2_gathersi<mode>"
>  (define_expand "avx2_gatherdi<mode>"
>   [(parallel [(set (match_operand:VEC_GATHER_MODE 0 "register_operand" "")
>                   (unspec:VEC_GATHER_MODE
> -                    [(match_operand:VEC_GATHER_MODE 1 "register_operand" "")
> +                    [(match_operand:<VEC_GATHER_SRCDI> 1 "register_operand" "")
>                      (mem:<ssescalarmode>
>                        (match_par_dup 7
>                          [(match_operand 2 "vsib_address_operand" "")
> -                          (match_operand:<AVXMODE48P_DI> 3 "register_operand" "")
> +                          (match_operand:<VEC_GATHER_IDXDI>
> +                             3 "register_operand" "")
>                           (match_operand:SI 5 "const1248_operand " "")]))
>                      (mem:BLK (scratch))
> -                     (match_operand:VEC_GATHER_MODE 4 "register_operand" "")]
> +                     (match_operand:<VEC_GATHER_SRCDI>
> +                       4 "register_operand" "")]
>                     UNSPEC_GATHER))
>              (clobber (match_scratch:VEC_GATHER_MODE 6 ""))])]
>   "TARGET_AVX2"
> @@ -12583,63 +12588,21 @@ (define_expand "avx2_gatherdi<mode>"
>  })
>
>  (define_insn "*avx2_gatherdi<mode>"
> -  [(set (match_operand:AVXMODE48P_DI 0 "register_operand" "=&x")
> -       (unspec:AVXMODE48P_DI
> -         [(match_operand:AVXMODE48P_DI 2 "register_operand" "0")
> +  [(set (match_operand:VEC_GATHER_MODE 0 "register_operand" "=&x")
> +       (unspec:VEC_GATHER_MODE
> +         [(match_operand:<VEC_GATHER_SRCDI> 2 "register_operand" "0")
>           (match_operator:<ssescalarmode> 7 "vsib_mem_operator"
>             [(unspec:P
>                [(match_operand:P 3 "vsib_address_operand" "p")
> -                (match_operand:<AVXMODE48P_DI> 4 "register_operand" "x")
> +                (match_operand:<VEC_GATHER_IDXDI> 4 "register_operand" "x")
>                 (match_operand:SI 6 "const1248_operand" "n")]
>                UNSPEC_VSIBADDR)])
>           (mem:BLK (scratch))
> -          (match_operand:AVXMODE48P_DI 5 "register_operand" "1")]
> +          (match_operand:<VEC_GATHER_SRCDI> 5 "register_operand" "1")]
>          UNSPEC_GATHER))
> -   (clobber (match_scratch:AVXMODE48P_DI 1 "=&x"))]
> -  "TARGET_AVX2"
> -  "v<sseintprefix>gatherq<ssemodesuffix>\t{%1, %7, %0|%0, %7, %1}"
> -  [(set_attr "type" "ssemov")
> -   (set_attr "prefix" "vex")
> -   (set_attr "mode" "<sseinsnmode>")])
> -
> -;; Special handling for VEX.256 with float arguments
> -;; since there're still xmms as operands
> -(define_expand "avx2_gatherdi<mode>256"
> -  [(parallel [(set (match_operand:VI4F_128 0 "register_operand" "")
> -                  (unspec:VI4F_128
> -                    [(match_operand:VI4F_128 1 "register_operand" "")
> -                     (mem:<ssescalarmode>
> -                       (match_par_dup 7
> -                         [(match_operand 2 "vsib_address_operand" "")
> -                          (match_operand:V4DI 3 "register_operand" "")
> -                          (match_operand:SI 5 "const1248_operand " "")]))
> -                     (mem:BLK (scratch))
> -                     (match_operand:VI4F_128 4 "register_operand" "")]
> -                    UNSPEC_GATHER))
> -             (clobber (match_scratch:VI4F_128 6 ""))])]
> -  "TARGET_AVX2"
> -{
> -  operands[7]
> -    = gen_rtx_UNSPEC (Pmode, gen_rtvec (3, operands[2], operands[3],
> -                                       operands[5]), UNSPEC_VSIBADDR);
> -})
> -
> -(define_insn "*avx2_gatherdi<mode>256"
> -  [(set (match_operand:VI4F_128 0 "register_operand" "=x")
> -       (unspec:VI4F_128
> -         [(match_operand:VI4F_128 2 "register_operand" "0")
> -          (match_operator:<ssescalarmode> 7 "vsib_mem_operator"
> -            [(unspec:P
> -               [(match_operand:P 3 "vsib_address_operand" "p")
> -                (match_operand:V4DI 4 "register_operand" "x")
> -                (match_operand:SI 6 "const1248_operand" "n")]
> -               UNSPEC_VSIBADDR)])
> -          (mem:BLK (scratch))
> -          (match_operand:VI4F_128 5 "register_operand" "1")]
> -         UNSPEC_GATHER))
> -   (clobber (match_scratch:VI4F_128 1 "=&x"))]
> +   (clobber (match_scratch:VEC_GATHER_MODE 1 "=&x"))]
>   "TARGET_AVX2"
> -  "v<sseintprefix>gatherq<ssemodesuffix>\t{%1, %7, %0|%0, %7, %1}"
> +  "v<sseintprefix>gatherq<ssemodesuffix>\t{%5, %7, %2|%2, %7, %5}"
>   [(set_attr "type" "ssemov")
>    (set_attr "prefix" "vex")
>    (set_attr "mode" "<sseinsnmode>")])
> --- gcc/config/i386/i386.c.jj   2011-11-04 08:52:19.000000000 +0100
> +++ gcc/config/i386/i386.c      2011-11-04 12:48:16.000000000 +0100
> @@ -25105,6 +25105,13 @@ enum ix86_builtins
>   IX86_BUILTIN_GATHERDIV4SI,
>   IX86_BUILTIN_GATHERDIV8SI,
>
> +  /* Alternate 4 element gather for the vectorizer where
> +     all operands are 32-byte wide.  */
> +  IX86_BUILTIN_GATHERALTSIV4DF,
> +  IX86_BUILTIN_GATHERALTDIV8SF,
> +  IX86_BUILTIN_GATHERALTSIV4DI,
> +  IX86_BUILTIN_GATHERALTDIV8SI,
> +
>   /* TFmode support builtins.  */
>   IX86_BUILTIN_INFQ,
>   IX86_BUILTIN_HUGE_VALQ,
> @@ -26883,6 +26890,22 @@ ix86_init_mmx_sse_builtins (void)
>               V4SI_FTYPE_V4SI_PCINT_V4DI_V4SI_INT,
>               IX86_BUILTIN_GATHERDIV8SI);
>
> +  def_builtin (OPTION_MASK_ISA_AVX2, "__builtin_ia32_gatheraltsiv4df ",
> +              V4DF_FTYPE_V4DF_PCDOUBLE_V8SI_V4DF_INT,
> +              IX86_BUILTIN_GATHERALTSIV4DF);
> +
> +  def_builtin (OPTION_MASK_ISA_AVX2, "__builtin_ia32_gatheraltdiv4sf256 ",
> +              V8SF_FTYPE_V8SF_PCFLOAT_V4DI_V8SF_INT,
> +              IX86_BUILTIN_GATHERALTDIV8SF);
> +
> +  def_builtin (OPTION_MASK_ISA_AVX2, "__builtin_ia32_gatheraltsiv4di ",
> +              V4DI_FTYPE_V4DI_PCINT64_V8SI_V4DI_INT,
> +              IX86_BUILTIN_GATHERALTSIV4DI);
> +
> +  def_builtin (OPTION_MASK_ISA_AVX2, "__builtin_ia32_gatheraltdiv4si256 ",
> +              V8SI_FTYPE_V8SI_PCINT_V4DI_V8SI_INT,
> +              IX86_BUILTIN_GATHERALTDIV8SI);
> +
>   /* MMX access to the vec_init patterns.  */
>   def_builtin_const (OPTION_MASK_ISA_MMX, "__builtin_ia32_vec_init_v2si",
>                     V2SI_FTYPE_INT_INT, IX86_BUILTIN_VEC_INIT_V2SI);
> @@ -28869,7 +28892,7 @@ rdrand_step:
>       icode = CODE_FOR_avx2_gatherdiv4sf;
>       goto gather_gen;
>     case IX86_BUILTIN_GATHERDIV8SF:
> -      icode = CODE_FOR_avx2_gatherdiv4sf256;
> +      icode = CODE_FOR_avx2_gatherdiv8sf;
>       goto gather_gen;
>     case IX86_BUILTIN_GATHERSIV2DI:
>       icode = CODE_FOR_avx2_gathersiv2di;
> @@ -28893,7 +28916,20 @@ rdrand_step:
>       icode = CODE_FOR_avx2_gatherdiv4si;
>       goto gather_gen;
>     case IX86_BUILTIN_GATHERDIV8SI:
> -      icode = CODE_FOR_avx2_gatherdiv4si256;
> +      icode = CODE_FOR_avx2_gatherdiv8si;
> +      goto gather_gen;
> +    case IX86_BUILTIN_GATHERALTSIV4DF:
> +      icode = CODE_FOR_avx2_gathersiv4df;
> +      goto gather_gen;
> +    case IX86_BUILTIN_GATHERALTDIV8SF:
> +      icode = CODE_FOR_avx2_gatherdiv8sf;
> +      goto gather_gen;
> +    case IX86_BUILTIN_GATHERALTSIV4DI:
> +      icode = CODE_FOR_avx2_gathersiv4df;
> +      goto gather_gen;
> +    case IX86_BUILTIN_GATHERALTDIV8SI:
> +      icode = CODE_FOR_avx2_gatherdiv8si;
> +      goto gather_gen;
>
>     gather_gen:
>       arg0 = CALL_EXPR_ARG (exp, 0);
> @@ -28912,8 +28948,39 @@ rdrand_step:
>       mode3 = insn_data[icode].operand[4].mode;
>       mode4 = insn_data[icode].operand[5].mode;
>
> -      if (target == NULL_RTX)
> -       target = gen_reg_rtx (insn_data[icode].operand[0].mode);
> +      if (target == NULL_RTX
> +         || GET_MODE (target) != insn_data[icode].operand[0].mode)
> +       subtarget = gen_reg_rtx (insn_data[icode].operand[0].mode);
> +      else
> +       subtarget = target;
> +
> +      if (fcode == IX86_BUILTIN_GATHERALTSIV4DF
> +         || fcode == IX86_BUILTIN_GATHERALTSIV4DI)
> +       {
> +         rtx half = gen_reg_rtx (V4SImode);
> +         if (!nonimmediate_operand (op2, V8SImode))
> +           op2 = copy_to_mode_reg (V8SImode, op2);
> +         emit_insn (gen_vec_extract_lo_v8si (half, op2));
> +         op2 = half;
> +       }
> +      else if (fcode == IX86_BUILTIN_GATHERALTDIV8SF
> +              || fcode == IX86_BUILTIN_GATHERALTDIV8SI)
> +       {
> +         rtx (*gen) (rtx, rtx);
> +         rtx half = gen_reg_rtx (mode0);
> +         if (mode0 == V4SFmode)
> +           gen = gen_vec_extract_lo_v8sf;
> +         else
> +           gen = gen_vec_extract_lo_v8si;
> +         if (!nonimmediate_operand (op0, GET_MODE (op0)))
> +           op0 = copy_to_mode_reg (GET_MODE (op0), op0);
> +         emit_insn (gen (half, op0));
> +         op0 = half;
> +         if (!nonimmediate_operand (op3, GET_MODE (op3)))
> +           op3 = copy_to_mode_reg (GET_MODE (op3), op3);
> +         emit_insn (gen (half, op3));
> +         op3 = half;
> +       }
>
>       /* Force memory operand only with base register here.  But we
>         don't want to do it on memory operand for other builtin
> @@ -28935,10 +29002,26 @@ rdrand_step:
>           error ("last argument must be scale 1, 2, 4, 8");
>           return const0_rtx;
>        }
> -      pat = GEN_FCN (icode) (target, op0, op1, op2, op3, op4);
> +      pat = GEN_FCN (icode) (subtarget, op0, op1, op2, op3, op4);
>       if (! pat)
>        return const0_rtx;
>       emit_insn (pat);
> +
> +      if (fcode == IX86_BUILTIN_GATHERDIV8SF
> +         || fcode == IX86_BUILTIN_GATHERDIV8SI)
> +       {
> +         enum machine_mode tmode = GET_MODE (subtarget) == V8SFmode
> +                                   ? V4SFmode : V4SImode;
> +         if (target == NULL_RTX)
> +           target = gen_reg_rtx (tmode);
> +         if (tmode == V4SFmode)
> +           emit_insn (gen_vec_extract_lo_v8sf (target, subtarget));
> +         else
> +           emit_insn (gen_vec_extract_lo_v8si (target, subtarget));
> +       }
> +      else
> +       target = subtarget;
> +
>       return target;
>
>     default:
> @@ -29443,6 +29526,73 @@ ix86_veclibabi_acml (enum built_in_funct
>   return new_fndecl;
>  }
>
> +/* Returns a decl of a function that implements gather load with
> +   memory type MEM_VECTYPE and index type INDEX_VECTYPE and SCALE.
> +   Return NULL_TREE if it is not available.  */
> +
> +static tree
> +ix86_vectorize_builtin_gather (const_tree mem_vectype,
> +                              const_tree index_type, int scale)
> +{
> +  bool si;
> +  enum ix86_builtins code;
> +
> +  if (! TARGET_AVX2)
> +    return NULL_TREE;
> +
> +  if ((TREE_CODE (index_type) != INTEGER_TYPE
> +       && !POINTER_TYPE_P (index_type))
> +      || (TYPE_MODE (index_type) != SImode
> +         && TYPE_MODE (index_type) != DImode))
> +    return NULL_TREE;
> +
> +  if (TYPE_PRECISION (index_type) > POINTER_SIZE)
> +    return NULL_TREE;
> +
> +  /* v*gather* insn sign extends index to pointer mode.  */
> +  if (TYPE_PRECISION (index_type) < POINTER_SIZE
> +      && TYPE_UNSIGNED (index_type))
> +    return NULL_TREE;
> +
> +  if (scale <= 0
> +      || scale > 8
> +      || (scale & (scale - 1)) != 0)
> +    return NULL_TREE;
> +
> +  si = TYPE_MODE (index_type) == SImode;
> +  switch (TYPE_MODE (mem_vectype))
> +    {
> +    case V2DFmode:
> +      code = si ? IX86_BUILTIN_GATHERSIV2DF : IX86_BUILTIN_GATHERDIV2DF;
> +      break;
> +    case V4DFmode:
> +      code = si ? IX86_BUILTIN_GATHERALTSIV4DF : IX86_BUILTIN_GATHERDIV4DF;
> +      break;
> +    case V2DImode:
> +      code = si ? IX86_BUILTIN_GATHERSIV2DI : IX86_BUILTIN_GATHERDIV2DI;
> +      break;
> +    case V4DImode:
> +      code = si ? IX86_BUILTIN_GATHERALTSIV4DI : IX86_BUILTIN_GATHERDIV4DI;
> +      break;
> +    case V4SFmode:
> +      code = si ? IX86_BUILTIN_GATHERSIV4SF : IX86_BUILTIN_GATHERDIV4SF;
> +      break;
> +    case V8SFmode:
> +      code = si ? IX86_BUILTIN_GATHERSIV8SF : IX86_BUILTIN_GATHERALTDIV8SF;
> +      break;
> +    case V4SImode:
> +      code = si ? IX86_BUILTIN_GATHERSIV4SI : IX86_BUILTIN_GATHERDIV4SI;
> +      break;
> +    case V8SImode:
> +      code = si ? IX86_BUILTIN_GATHERSIV8SI : IX86_BUILTIN_GATHERALTDIV8SI;
> +      break;
> +    default:
> +      return NULL_TREE;
> +    }
> +
> +  return ix86_builtins[code];
> +}
> +
>  /* Returns a code for a target-specific builtin that implements
>    reciprocal of the function, or NULL_TREE if not available.  */
>
> @@ -37642,6 +37792,9 @@ ix86_autovectorize_vector_sizes (void)
>  #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION \
>   ix86_builtin_vectorized_function
>
> +#undef TARGET_VECTORIZE_BUILTIN_GATHER
> +#define TARGET_VECTORIZE_BUILTIN_GATHER ix86_vectorize_builtin_gather
> +
>  #undef TARGET_BUILTIN_RECIPROCAL
>  #define TARGET_BUILTIN_RECIPROCAL ix86_builtin_reciprocal
>
> --- gcc/testsuite/gcc.target/i386/avx2-gather-1.c.jj    2011-11-04 08:53:13.000000000 +0100
> +++ gcc/testsuite/gcc.target/i386/avx2-gather-1.c       2011-11-04 08:53:13.000000000 +0100
> @@ -0,0 +1,215 @@
> +/* { dg-do run } */
> +/* { dg-require-effective-target avx2 } */
> +/* { dg-options "-O3 -mavx2" } */
> +
> +#include "avx2-check.h"
> +
> +#define N 1024
> +float vf1[N+16], vf2[N];
> +double vd1[N+16], vd2[N];
> +int k[N];
> +long l[N];
> +short n[N];
> +
> +__attribute__((noinline, noclone)) void
> +f1 (void)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    vf2[i] = vf1[k[i]];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f2 (void)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    n[i] = (int) vf1[k[i]];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f3 (int x)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    vf2[i] = vf1[k[i] + x];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f4 (int x)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    n[i] = (int) vf1[k[i] + x];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f5 (void)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    vd2[i] = vd1[k[i]];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f6 (void)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    n[i] = (int) vd1[k[i]];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f7 (int x)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    vd2[i] = vd1[k[i] + x];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f8 (int x)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    n[i] = (int) vd1[k[i] + x];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f9 (void)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    vf2[i] = vf1[l[i]];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f10 (void)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    n[i] = (int) vf1[l[i]];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f11 (long x)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    vf2[i] = vf1[l[i] + x];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f12 (long x)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    n[i] = (int) vf1[l[i] + x];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f13 (void)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    vd2[i] = vd1[l[i]];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f14 (void)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    n[i] = (int) vd1[l[i]];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f15 (long x)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    vd2[i] = vd1[l[i] + x];
> +}
> +
> +__attribute__((noinline, noclone)) void
> +f16 (long x)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    n[i] = (int) vd1[l[i] + x];
> +}
> +
> +static void
> +avx2_test (void)
> +{
> +  int i;
> +
> +  for (i = 0; i < N + 16; i++)
> +    {
> +      asm ("");
> +      vf1[i] = 17.0f + i;
> +      vd1[i] = 19.0 + i;
> +    }
> +  for (i = 0; i < N; i++)
> +    {
> +      asm ("");
> +      k[i] = (i * 731) & (N - 1);
> +      l[i] = (i * 657) & (N - 1);
> +    }
> +
> +  f1 ();
> +  f2 ();
> +  for (i = 0; i < N; i++)
> +    if (vf2[i] != ((i * 731) & (N - 1)) + 17
> +       || n[i] != ((i * 731) & (N - 1)) + 17)
> +      abort ();
> +
> +  f3 (12);
> +  f4 (14);
> +  for (i = 0; i < N; i++)
> +    if (vf2[i] != ((i * 731) & (N - 1)) + 17 + 12
> +       || n[i] != ((i * 731) & (N - 1)) + 17 + 14)
> +      abort ();
> +
> +  f5 ();
> +  f6 ();
> +  for (i = 0; i < N; i++)
> +    if (vd2[i] != ((i * 731) & (N - 1)) + 19
> +       || n[i] != ((i * 731) & (N - 1)) + 19)
> +      abort ();
> +
> +  f7 (7);
> +  f8 (9);
> +  for (i = 0; i < N; i++)
> +    if (vd2[i] != ((i * 731) & (N - 1)) + 19 + 7
> +       || n[i] != ((i * 731) & (N - 1)) + 19 + 9)
> +      abort ();
> +
> +  f9 ();
> +  f10 ();
> +  for (i = 0; i < N; i++)
> +    if (vf2[i] != ((i * 657) & (N - 1)) + 17
> +       || n[i] != ((i * 657) & (N - 1)) + 17)
> +      abort ();
> +
> +  f11 (2);
> +  f12 (4);
> +  for (i = 0; i < N; i++)
> +    if (vf2[i] != ((i * 657) & (N - 1)) + 17 + 2
> +       || n[i] != ((i * 657) & (N - 1)) + 17 + 4)
> +      abort ();
> +
> +  f13 ();
> +  f14 ();
> +  for (i = 0; i < N; i++)
> +    if (vd2[i] != ((i * 657) & (N - 1)) + 19
> +       || n[i] != ((i * 657) & (N - 1)) + 19)
> +      abort ();
> +
> +  f15 (13);
> +  f16 (15);
> +  for (i = 0; i < N; i++)
> +    if (vd2[i] != ((i * 657) & (N - 1)) + 19 + 13
> +       || n[i] != ((i * 657) & (N - 1)) + 19 + 15)
> +      abort ();
> +}
> --- gcc/testsuite/gcc.target/i386/avx2-gather-2.c.jj    2011-11-04 08:53:13.000000000 +0100
> +++ gcc/testsuite/gcc.target/i386/avx2-gather-2.c       2011-11-04 08:53:13.000000000 +0100
> @@ -0,0 +1,7 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O3 -mavx2 -fdump-tree-vect-details" } */
> +
> +#include "avx2-gather-1.c"
> +
> +/* { dg-final { scan-tree-dump-times "note: vectorized 1 loops in function" 16 "vect" } } */
> +/* { dg-final { cleanup-tree-dump "vect" } } */
> --- gcc/testsuite/gcc.target/i386/avx2-gather-3.c.jj    2011-11-04 08:53:13.000000000 +0100
> +++ gcc/testsuite/gcc.target/i386/avx2-gather-3.c       2011-11-04 08:53:13.000000000 +0100
> @@ -0,0 +1,167 @@
> +/* { dg-do run } */
> +/* { dg-require-effective-target avx2 } */
> +/* { dg-options "-O3 -mavx2 -ffast-math" } */
> +
> +#include "avx2-check.h"
> +
> +#define N 1024
> +float f[N];
> +double d[N];
> +int k[N];
> +float *l[N];
> +double *n[N];
> +int **m[N];
> +long **o[N];
> +long q[N];
> +long *r[N];
> +int *s[N];
> +
> +__attribute__((noinline, noclone)) float
> +f1 (void)
> +{
> +  int i;
> +  float g = 0.0;
> +  for (i = 0; i < N / 2; i++)
> +    g += f[k[i]];
> +  return g;
> +}
> +
> +__attribute__((noinline, noclone)) float
> +f2 (float *p)
> +{
> +  int i;
> +  float g = 0.0;
> +  for (i = 0; i < N / 2; i++)
> +    g += p[k[i]];
> +  return g;
> +}
> +
> +__attribute__((noinline, noclone)) float
> +f3 (void)
> +{
> +  int i;
> +  float g = 0.0;
> +  for (i = 0; i < N / 2; i++)
> +    g += *l[i];
> +  return g;
> +}
> +
> +__attribute__((noinline, noclone)) int
> +f4 (void)
> +{
> +  int i;
> +  int g = 0;
> +  for (i = 0; i < N / 2; i++)
> +    g += **m[i];
> +  return g;
> +}
> +
> +__attribute__((noinline, noclone)) double
> +f5 (void)
> +{
> +  int i;
> +  double g = 0.0;
> +  for (i = 0; i < N / 2; i++)
> +    g += d[k[i]];
> +  return g;
> +}
> +
> +__attribute__((noinline, noclone)) double
> +f6 (double *p)
> +{
> +  int i;
> +  double g = 0.0;
> +  for (i = 0; i < N / 2; i++)
> +    g += p[k[i]];
> +  return g;
> +}
> +
> +__attribute__((noinline, noclone)) double
> +f7 (void)
> +{
> +  int i;
> +  double g = 0.0;
> +  for (i = 0; i < N / 2; i++)
> +    g += *n[i];
> +  return g;
> +}
> +
> +__attribute__((noinline, noclone)) int
> +f8 (void)
> +{
> +  int i;
> +  int g = 0;
> +  for (i = 0; i < N / 2; i++)
> +    g += **o[i];
> +  return g;
> +}
> +
> +__attribute__((noinline, noclone)) float
> +f9 (void)
> +{
> +  int i;
> +  float g = 0.0;
> +  for (i = 0; i < N / 2; i++)
> +    g += f[q[i]];
> +  return g;
> +}
> +
> +__attribute__((noinline, noclone)) float
> +f10 (float *p)
> +{
> +  int i;
> +  float g = 0.0;
> +  for (i = 0; i < N / 2; i++)
> +    g += p[q[i]];
> +  return g;
> +}
> +
> +__attribute__((noinline, noclone)) double
> +f11 (void)
> +{
> +  int i;
> +  double g = 0.0;
> +  for (i = 0; i < N / 2; i++)
> +    g += d[q[i]];
> +  return g;
> +}
> +
> +__attribute__((noinline, noclone)) double
> +f12 (double *p)
> +{
> +  int i;
> +  double g = 0.0;
> +  for (i = 0; i < N / 2; i++)
> +    g += p[q[i]];
> +  return g;
> +}
> +
> +static void
> +avx2_test (void)
> +{
> +  int i;
> +
> +  for (i = 0; i < N; i++)
> +    {
> +      asm ("");
> +      f[i] = -256.0f + i;
> +      d[i] = -258.0 + i;
> +      k[i] = (i * 731) & (N - 1);
> +      q[i] = (i * 657) & (N - 1);
> +      l[i] = &f[(i * 239) & (N - 1)];
> +      n[i] = &d[(i * 271) & (N - 1)];
> +      r[i] = &q[(i * 323) & (N - 1)];
> +      s[i] = &k[(i * 565) & (N - 1)];
> +      m[i] = &s[(i * 13) & (N - 1)];
> +      o[i] = &r[(i * 19) & (N - 1)];
> +    }
> +
> +  if (f1 () != 136448.0f || f2 (f) != 136448.0f || f3 () != 130304.0)
> +    abort ();
> +  if (f4 () != 261376 || f5 () != 135424.0 || f6 (d) != 135424.0)
> +    abort ();
> +  if (f7 () != 129280.0 || f8 () != 259840L || f9 () != 130816.0f)
> +    abort ();
> +  if (f10 (f) != 130816.0f || f11 () != 129792.0 || f12 (d) != 129792.0)
> +    abort ();
> +}
> --- gcc/testsuite/gcc.target/i386/avx2-gather-4.c.jj    2011-11-04 08:54:11.000000000 +0100
> +++ gcc/testsuite/gcc.target/i386/avx2-gather-4.c       2011-11-04 08:54:11.000000000 +0100
> @@ -0,0 +1,38 @@
> +/* { dg-do run } */
> +/* { dg-require-effective-target avx2 } */
> +/* { dg-options "-O3 -mavx2" } */
> +
> +#include "avx2-check.h"
> +
> +#define N 1024
> +int a[N], b[N], c[N], d[N];
> +
> +__attribute__((noinline, noclone)) void
> +foo (float *__restrict p, float *__restrict q, float *__restrict r,
> +     long s1, long s2, long s3)
> +{
> +  int i;
> +  for (i = 0; i < N; i++)
> +    p[i] = q[a[i] * s1 + b[i] * s2 + s3] * r[c[i] * s1 + d[i] * s2 + s3];
> +}
> +
> +static void
> +avx2_test (void)
> +{
> +  int i;
> +  float e[N], f[N], g[N];
> +  for (i = 0; i < N; i++)
> +    {
> +      a[i] = (i * 7) & (N / 8 - 1);
> +      b[i] = (i * 13) & (N / 8 - 1);
> +      c[i] = (i * 23) & (N / 8 - 1);
> +      d[i] = (i * 5) & (N / 8 - 1);
> +      e[i] = 16.5 + i;
> +      f[i] = 127.5 - i;
> +    }
> +  foo (g, e, f, 3, 2, 4);
> +  for (i = 0; i < N; i++)
> +    if (g[i] != (float) ((20.5 + a[i] * 3 + b[i] * 2)
> +                        * (123.5 - c[i] * 3 - d[i] * 2)))
> +      abort ();
> +}
>
>
>        Jakub
>

  reply	other threads:[~2011-11-04 22:26 UTC|newest]

Thread overview: 8+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2011-11-02 20:19 Patch ping Jakub Jelinek
2011-11-04 10:11 ` Richard Guenther
2011-11-04 10:39   ` Jakub Jelinek
2011-11-04 11:44     ` Richard Guenther
2011-11-04 13:32       ` [PATCH] Gather vectorization (PR tree-optimization/50789, take 2) Jakub Jelinek
2011-11-04 23:54         ` Richard Guenther [this message]
2011-11-07 15:49         ` Uros Bizjak
2011-11-04 14:09       ` Patch ping Michael Matz

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CAFiYyc1Gg85Rb5RJdZkO+e-P+4VC1C3YOSetsPBF-nSaXyaYbg@mail.gmail.com \
    --to=richard.guenther@gmail.com \
    --cc=gcc-patches@gcc.gnu.org \
    --cc=ira.rosen@linaro.org \
    --cc=jakub@redhat.com \
    --cc=rguenther@suse.de \
    --cc=rth@redhat.com \
    --cc=ubizjak@gmail.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).