Re: Re: [PATCH V3] VECT: Apply LEN_MASK_GATHER_LOAD/SCATTER_STORE into vectorizer

["juzhe.zhong@rivai.ai" <juzhe.zhong@rivai.ai>]

[-- Attachment #1: Type: text/plain, Size: 18907 bytes --]

Hi, Richi.

>> Eh, it's obvious that we should have the "vectorized" form
>> also in the 'scalar' variant.  If you think there's no reasonable
>> way to add a value for len or bias then instead re-order the
>> arguments so 'mask' comes first and the len/bias pair last.

I found this patch is buggy when I am testing LEN_MAS_GATHER_LOAD.
And reoder 'mask' comes first can not help.

Here, we have 'GATHER_LOAD' and 'MASK_GATHER_LOAD',
For GATHER_LOAD ((sizetype) src_18(D), _6, 1, 0);
I change it into:

LEN_MASK_GATHER_LOAD ((sizetype) src_18(D), _6, 1, 0);
In this situation,  internal_fn_mask_index 
should return -1.

Wheras, MASK_GATHER_LOAD ((sizetype) _56, _8, 1, 0, _33);
I change it into LEN_MASK_GATHER_LOAD ((sizetype) _56, _8, 1, 0, _33);
it should return the index = 4.

I can't differentiate them by only using LEN_MASK_GATHER_LOAD.
Could I revise internal_fn_mask_index 
as follows ?

int
internal_fn_mask_index (internal_fn fn, int nargs)
{
  switch (fn)
    {
    case IFN_MASK_LOAD:
    case IFN_MASK_LOAD_LANES:
    case IFN_MASK_STORE:
    case IFN_MASK_STORE_LANES:
      return 2;

    case IFN_MASK_GATHER_LOAD:
    case IFN_MASK_SCATTER_STORE:
    case IFN_LEN_MASK_LOAD:
    case IFN_LEN_MASK_STORE:
      return 4;

    /* LEN_MASK_GATHER_LOAD/LEN_MASK_SCATTER_STORE are different
       on arguments before and after vectorized.

         Before vectorized:
           LEN_MASK_GATHER_LOAD (ptr, align, offset, mask);

         After vectorized:
           LEN_MASK_GATHER_LOAD (ptr, align, offset, len, bias, mask);
    */
    case IFN_LEN_MASK_GATHER_LOAD:
    case IFN_LEN_MASK_SCATTER_STORE:
      return nargs == 4 ? -1 : nargs == 5 ? 4 : 6;

    default:
      return (conditional_internal_fn_code (fn) != ERROR_MARK
              || get_unconditional_internal_fn (fn) != IFN_LAST ? 0 : -1);
    }
}


Thanks.


juzhe.zhong@rivai.ai
 
From: Richard Biener
Date: 2023-07-04 19:05
To: Ju-Zhe Zhong
CC: gcc-patches; richard.sandiford
Subject: Re: [PATCH V3] VECT: Apply LEN_MASK_GATHER_LOAD/SCATTER_STORE into vectorizer
On Tue, 4 Jul 2023, juzhe.zhong@rivai.ai wrote:
 
> From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
> 
> Hi, Richard and Richi.
> 
> Address comments from Richard.
> 
> Make gs_info.ifn = LEN_MASK_GATHER_LOAD/LEN_MASK_SCATTER_STORE.
> 
> Since:
>     /* LEN_MASK_GATHER_LOAD/LEN_MASK_SCATTER_STORE are different
>        on arguments before and after vectorized.
> 
> Before vectorized:
>    LEN_MASK_GATHER_LOAD (ptr, align, offset, mask);
> 
> After vectorized:
>    LEN_MASK_GATHER_LOAD (ptr, align, offset, len, bias, mask);
>     */
> 
> I add "vectorized_p" default argument into internal_fn_mask_index.
> So that we could simplify the codes.
 
Eh, it's obvious that we should have the "vectorized" form
also in the 'scalar' variant.  If you think there's no reasonable
way to add a value for len or bias then instead re-order the
arguments so 'mask' comes first and the len/bias pair last.
 
But IMHO "any" len/bias value should do here.
 
The rest looks OK now.
 
Thanks,
Richard.
 
> The len_mask_gather_load/len_mask_scatter_store patterns have been added.
> Now, this patch applies them into vectorizer.
> 
> Here is the example:
> 
> void
> f (int *restrict a,
>    int *restrict b, int n,
>    int base, int step,
>    int *restrict cond)
> {
>   for (int i = 0; i < n; ++i)
>     {
>       if (cond[i])
>         a[i * 4] = b[i];
>     }
> }
> 
> Gimple IR:
> 
>   <bb 3> [local count: 105119324]:
>   _58 = (unsigned long) n_13(D);
> 
>   <bb 4> [local count: 630715945]:
>   # vectp_cond.7_45 = PHI <vectp_cond.7_46(4), cond_14(D)(3)>
>   # vectp_b.11_51 = PHI <vectp_b.11_52(4), b_15(D)(3)>
>   # vectp_a.14_55 = PHI <vectp_a.14_56(4), a_16(D)(3)>
>   # ivtmp_59 = PHI <ivtmp_60(4), _58(3)>
>   _61 = .SELECT_VL (ivtmp_59, POLY_INT_CST [2, 2]);
>   ivtmp_44 = _61 * 4;
>   vect__4.9_47 = .LEN_MASK_LOAD (vectp_cond.7_45, 32B, _61, 0, { -1, ... });
>   mask__24.10_49 = vect__4.9_47 != { 0, ... };
>   vect__8.13_53 = .LEN_MASK_LOAD (vectp_b.11_51, 32B, _61, 0, mask__24.10_49);
>   ivtmp_54 = _61 * 16;
>   .LEN_MASK_SCATTER_STORE (vectp_a.14_55, { 0, 16, 32, ... }, 1, vect__8.13_53, _61, 0, mask__24.10_49);
>   vectp_cond.7_46 = vectp_cond.7_45 + ivtmp_44;
>   vectp_b.11_52 = vectp_b.11_51 + ivtmp_44;
>   vectp_a.14_56 = vectp_a.14_55 + ivtmp_54;
>   ivtmp_60 = ivtmp_59 - _61;
>   if (ivtmp_60 != 0)
>     goto <bb 4>; [83.33%]
>   else
>     goto <bb 5>; [16.67%]
> 
> gcc/ChangeLog:
> 
>         * internal-fn.cc (internal_fn_mask_index): Apply LEN_MASK_GATHER_LOAD/SCATTER_STORE into vectorizer.
>         * internal-fn.h (internal_fn_mask_index): Ditto.
>         * optabs-query.cc (supports_vec_gather_load_p): Ditto.
>         (supports_vec_scatter_store_p): Ditto.
>         * tree-vect-data-refs.cc (vect_gather_scatter_fn_p): Ditto.
>         * tree-vect-stmts.cc (exist_non_indexing_operands_for_use_p): Ditto.
>         (check_load_store_for_partial_vectors): Ditto.
>         (vect_get_strided_load_store_ops): Ditto.
>         (vectorizable_store): Ditto.
>         (vectorizable_load): Ditto.
> 
> ---
>  gcc/internal-fn.cc         |  16 ++++-
>  gcc/internal-fn.h          |   2 +-
>  gcc/optabs-query.cc        |   2 +
>  gcc/tree-vect-data-refs.cc |  18 ++++-
>  gcc/tree-vect-stmts.cc     | 135 +++++++++++++++++++++++++++++++------
>  5 files changed, 150 insertions(+), 23 deletions(-)
> 
> diff --git a/gcc/internal-fn.cc b/gcc/internal-fn.cc
> index 303df102d81..2c78c870de8 100644
> --- a/gcc/internal-fn.cc
> +++ b/gcc/internal-fn.cc
> @@ -4483,7 +4483,7 @@ internal_fn_len_index (internal_fn fn)
>     otherwise return -1.  */
>  
>  int
> -internal_fn_mask_index (internal_fn fn)
> +internal_fn_mask_index (internal_fn fn, bool vectoried_p)
>  {
>    switch (fn)
>      {
> @@ -4499,9 +4499,21 @@ internal_fn_mask_index (internal_fn fn)
>      case IFN_LEN_MASK_STORE:
>        return 4;
>  
> +    /* LEN_MASK_GATHER_LOAD/LEN_MASK_SCATTER_STORE are different
> +       on arguments before and after vectorized.
> +
> + Before vectorized:
> +    LEN_MASK_GATHER_LOAD (ptr, align, offset, mask);
> +
> + After vectorized:
> +    LEN_MASK_GATHER_LOAD (ptr, align, offset, len, bias, mask);
> +    */
>      case IFN_LEN_MASK_GATHER_LOAD:
>      case IFN_LEN_MASK_SCATTER_STORE:
> -      return 6;
> +      if (vectoried_p)
> + return 6;
> +      else
> + return 4;
>  
>      default:
>        return (conditional_internal_fn_code (fn) != ERROR_MARK
> diff --git a/gcc/internal-fn.h b/gcc/internal-fn.h
> index 4234bbfed87..e9168c16297 100644
> --- a/gcc/internal-fn.h
> +++ b/gcc/internal-fn.h
> @@ -233,7 +233,7 @@ extern bool can_interpret_as_conditional_op_p (gimple *, tree *,
>  extern bool internal_load_fn_p (internal_fn);
>  extern bool internal_store_fn_p (internal_fn);
>  extern bool internal_gather_scatter_fn_p (internal_fn);
> -extern int internal_fn_mask_index (internal_fn);
> +extern int internal_fn_mask_index (internal_fn, bool = true);
>  extern int internal_fn_len_index (internal_fn);
>  extern int internal_fn_stored_value_index (internal_fn);
>  extern bool internal_gather_scatter_fn_supported_p (internal_fn, tree,
> diff --git a/gcc/optabs-query.cc b/gcc/optabs-query.cc
> index 2fdd0d34354..bf1f484e874 100644
> --- a/gcc/optabs-query.cc
> +++ b/gcc/optabs-query.cc
> @@ -676,6 +676,7 @@ supports_vec_gather_load_p (machine_mode mode)
>      this_fn_optabs->supports_vec_gather_load[mode]
>        = (supports_vec_convert_optab_p (gather_load_optab, mode)
>  || supports_vec_convert_optab_p (mask_gather_load_optab, mode)
> + || supports_vec_convert_optab_p (len_mask_gather_load_optab, mode)
>  ? 1 : -1);
>  
>    return this_fn_optabs->supports_vec_gather_load[mode] > 0;
> @@ -692,6 +693,7 @@ supports_vec_scatter_store_p (machine_mode mode)
>      this_fn_optabs->supports_vec_scatter_store[mode]
>        = (supports_vec_convert_optab_p (scatter_store_optab, mode)
>  || supports_vec_convert_optab_p (mask_scatter_store_optab, mode)
> + || supports_vec_convert_optab_p (len_mask_scatter_store_optab, mode)
>  ? 1 : -1);
>  
>    return this_fn_optabs->supports_vec_scatter_store[mode] > 0;
> diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
> index ebe93832b1e..ab2af103cb4 100644
> --- a/gcc/tree-vect-data-refs.cc
> +++ b/gcc/tree-vect-data-refs.cc
> @@ -3873,16 +3873,24 @@ vect_gather_scatter_fn_p (vec_info *vinfo, bool read_p, bool masked_p,
>      return false;
>  
>    /* Work out which function we need.  */
> -  internal_fn ifn, alt_ifn;
> +  internal_fn ifn, alt_ifn, alt_ifn2;
>    if (read_p)
>      {
>        ifn = masked_p ? IFN_MASK_GATHER_LOAD : IFN_GATHER_LOAD;
>        alt_ifn = IFN_MASK_GATHER_LOAD;
> +      /* When target supports LEN_MASK_GATHER_LOAD, we always
> + use LEN_MASK_GATHER_LOAD regardless whether len and
> + mask are valid or not.  */
> +      alt_ifn2 = IFN_LEN_MASK_GATHER_LOAD;
>      }
>    else
>      {
>        ifn = masked_p ? IFN_MASK_SCATTER_STORE : IFN_SCATTER_STORE;
>        alt_ifn = IFN_MASK_SCATTER_STORE;
> +      /* When target supports LEN_MASK_SCATTER_STORE, we always
> + use LEN_MASK_SCATTER_STORE regardless whether len and
> + mask are valid or not.  */
> +      alt_ifn2 = IFN_LEN_MASK_SCATTER_STORE;
>      }
>  
>    for (;;)
> @@ -3909,6 +3917,14 @@ vect_gather_scatter_fn_p (vec_info *vinfo, bool read_p, bool masked_p,
>    *offset_vectype_out = offset_vectype;
>    return true;
>  }
> +      else if (internal_gather_scatter_fn_supported_p (alt_ifn2, vectype,
> +        memory_type,
> +        offset_vectype, scale))
> + {
> +   *ifn_out = alt_ifn2;
> +   *offset_vectype_out = offset_vectype;
> +   return true;
> + }
>  
>        if (TYPE_PRECISION (offset_type) >= POINTER_SIZE
>    && TYPE_PRECISION (offset_type) >= element_bits)
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index a0c39268bf0..33ec33f8b8d 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -433,7 +433,7 @@ exist_non_indexing_operands_for_use_p (tree use, stmt_vec_info stmt_info)
>        if (call && gimple_call_internal_p (call))
>  {
>    internal_fn ifn = gimple_call_internal_fn (call);
> -   int mask_index = internal_fn_mask_index (ifn);
> +   int mask_index = internal_fn_mask_index (ifn, false);
>    if (mask_index >= 0
>        && use == gimple_call_arg (call, mask_index))
>      return true;
> @@ -1771,6 +1771,18 @@ check_load_store_for_partial_vectors (loop_vec_info loop_vinfo, tree vectype,
>     gs_info->offset_vectype,
>     gs_info->scale))
>  {
> +   ifn = (is_load
> + ? IFN_LEN_MASK_GATHER_LOAD
> + : IFN_LEN_MASK_SCATTER_STORE);
> +   if (internal_gather_scatter_fn_supported_p (ifn, vectype,
> +       gs_info->memory_type,
> +       gs_info->offset_vectype,
> +       gs_info->scale))
> +     {
> +       vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo);
> +       vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, 1);
> +       return;
> +     }
>    if (dump_enabled_p ())
>      dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
>       "can't operate on partial vectors because"
> @@ -3129,16 +3141,39 @@ vect_get_gather_scatter_ops (loop_vec_info loop_vinfo,
>  static void
>  vect_get_strided_load_store_ops (stmt_vec_info stmt_info,
>  loop_vec_info loop_vinfo,
> + gimple_stmt_iterator *gsi,
>  gather_scatter_info *gs_info,
> - tree *dataref_bump, tree *vec_offset)
> + tree *dataref_bump, tree *vec_offset,
> + vec_loop_lens *loop_lens)
>  {
>    struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
>    tree vectype = STMT_VINFO_VECTYPE (stmt_info);
>  
> -  tree bump = size_binop (MULT_EXPR,
> -   fold_convert (sizetype, unshare_expr (DR_STEP (dr))),
> -   size_int (TYPE_VECTOR_SUBPARTS (vectype)));
> -  *dataref_bump = cse_and_gimplify_to_preheader (loop_vinfo, bump);
> +  if (LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo))
> +    {
> +      /* _31 = .SELECT_VL (ivtmp_29, POLY_INT_CST [4, 4]);
> + ivtmp_8 = _31 * 16 (step in bytes);
> + .LEN_MASK_SCATTER_STORE (vectp_a.9_7, ... );
> + vectp_a.9_26 = vectp_a.9_7 + ivtmp_8;  */
> +      tree loop_len
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, vectype, 0, 0);
> +      tree tmp
> + = fold_build2 (MULT_EXPR, sizetype,
> +        fold_convert (sizetype, unshare_expr (DR_STEP (dr))),
> +        loop_len);
> +      tree bump = make_temp_ssa_name (sizetype, NULL, "ivtmp");
> +      gassign *assign = gimple_build_assign (bump, tmp);
> +      gsi_insert_before (gsi, assign, GSI_SAME_STMT);
> +      *dataref_bump = bump;
> +    }
> +  else
> +    {
> +      tree bump
> + = size_binop (MULT_EXPR,
> +       fold_convert (sizetype, unshare_expr (DR_STEP (dr))),
> +       size_int (TYPE_VECTOR_SUBPARTS (vectype)));
> +      *dataref_bump = cse_and_gimplify_to_preheader (loop_vinfo, bump);
> +    }
>  
>    /* The offset given in GS_INFO can have pointer type, so use the element
>       type of the vector instead.  */
> @@ -8013,7 +8048,7 @@ vectorizable_store (vec_info *vinfo,
>    return false;
>  }
>  
> -      int mask_index = internal_fn_mask_index (ifn);
> +      int mask_index = internal_fn_mask_index (ifn, false);
>        if (mask_index >= 0
>    && !vect_check_scalar_mask (vinfo, stmt_info, slp_node, mask_index,
>        &mask, NULL, &mask_dt, &mask_vectype))
> @@ -8685,8 +8720,8 @@ vectorizable_store (vec_info *vinfo,
>    else if (memory_access_type == VMAT_GATHER_SCATTER)
>      {
>        aggr_type = elem_type;
> -      vect_get_strided_load_store_ops (stmt_info, loop_vinfo, &gs_info,
> -        &bump, &vec_offset);
> +      vect_get_strided_load_store_ops (stmt_info, loop_vinfo, gsi, &gs_info,
> +        &bump, &vec_offset, loop_lens);
>      }
>    else
>      {
> @@ -8915,6 +8950,8 @@ vectorizable_store (vec_info *vinfo,
>        unsigned HOST_WIDE_INT align;
>  
>        tree final_mask = NULL_TREE;
> +       tree final_len = NULL_TREE;
> +       tree bias = NULL_TREE;
>        if (loop_masks)
>  final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
>  vec_num * ncopies,
> @@ -8929,8 +8966,41 @@ vectorizable_store (vec_info *vinfo,
>    if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
>      vec_offset = vec_offsets[vec_num * j + i];
>    tree scale = size_int (gs_info.scale);
> +
> +   if (gs_info.ifn == IFN_LEN_MASK_SCATTER_STORE)
> +     {
> +       if (loop_lens)
> + {
> +   final_len
> +     = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> + vec_num * j + i, 1);
> + }
> +       else
> + {
> +   tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> +   final_len
> +     = build_int_cst (iv_type,
> +      TYPE_VECTOR_SUBPARTS (vectype));
> + }
> +       signed char biasval
> + = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
> +       bias = build_int_cst (intQI_type_node, biasval);
> +       if (!final_mask)
> + {
> +   mask_vectype = truth_type_for (vectype);
> +   final_mask = build_minus_one_cst (mask_vectype);
> + }
> +     }
> +
>    gcall *call;
> -   if (final_mask)
> +   if (final_len && final_len)
> +     call
> +       = gimple_build_call_internal (IFN_LEN_MASK_SCATTER_STORE,
> +     7, dataref_ptr, vec_offset,
> +     scale, vec_oprnd, final_len,
> +     bias, final_mask);
> +   else if (final_mask)
>      call = gimple_build_call_internal
>        (IFN_MASK_SCATTER_STORE, 5, dataref_ptr, vec_offset,
>         scale, vec_oprnd, final_mask);
> @@ -9047,9 +9117,6 @@ vectorizable_store (vec_info *vinfo,
>        machine_mode vmode = TYPE_MODE (vectype);
>        machine_mode new_vmode = vmode;
>        internal_fn partial_ifn = IFN_LAST;
> -       /* Produce 'len' and 'bias' argument.  */
> -       tree final_len = NULL_TREE;
> -       tree bias = NULL_TREE;
>        if (loop_lens)
>  {
>    opt_machine_mode new_ovmode
> @@ -10177,8 +10244,8 @@ vectorizable_load (vec_info *vinfo,
>    else if (memory_access_type == VMAT_GATHER_SCATTER)
>      {
>        aggr_type = elem_type;
> -      vect_get_strided_load_store_ops (stmt_info, loop_vinfo, &gs_info,
> -        &bump, &vec_offset);
> +      vect_get_strided_load_store_ops (stmt_info, loop_vinfo, gsi, &gs_info,
> +        &bump, &vec_offset, loop_lens);
>      }
>    else
>      {
> @@ -10339,6 +10406,8 @@ vectorizable_load (vec_info *vinfo,
>    for (i = 0; i < vec_num; i++)
>      {
>        tree final_mask = NULL_TREE;
> +       tree final_len = NULL_TREE;
> +       tree bias = NULL_TREE;
>        if (loop_masks
>    && memory_access_type != VMAT_INVARIANT)
>  final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
> @@ -10368,8 +10437,39 @@ vectorizable_load (vec_info *vinfo,
>    vec_offset = vec_offsets[vec_num * j + i];
>  tree zero = build_zero_cst (vectype);
>  tree scale = size_int (gs_info.scale);
> +
> + if (gs_info.ifn == IFN_LEN_MASK_GATHER_LOAD)
> +   {
> +     if (loop_lens)
> +       {
> + final_len = vect_get_loop_len (
> +   loop_vinfo, gsi, loop_lens, vec_num * ncopies,
> +   vectype, vec_num * j + i, 1);
> +       }
> +     else
> +       {
> + tree iv_type
> +   = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + final_len = build_int_cst (
> +   iv_type, TYPE_VECTOR_SUBPARTS (vectype));
> +       }
> +     signed char biasval
> +       = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
> +     bias = build_int_cst (intQI_type_node, biasval);
> +     if (!final_mask)
> +       {
> + mask_vectype = truth_type_for (vectype);
> + final_mask = build_minus_one_cst (mask_vectype);
> +       }
> +   }
> +
>  gcall *call;
> - if (final_mask)
> + if (final_len && final_mask)
> +   call = gimple_build_call_internal (
> +     IFN_LEN_MASK_GATHER_LOAD, 7, dataref_ptr,
> +     vec_offset, scale, zero, final_len, bias,
> +     final_mask);
> + else if (final_mask)
>    call = gimple_build_call_internal
>      (IFN_MASK_GATHER_LOAD, 5, dataref_ptr,
>       vec_offset, scale, zero, final_mask);
> @@ -10462,9 +10562,6 @@ vectorizable_load (vec_info *vinfo,
>      machine_mode vmode = TYPE_MODE (vectype);
>      machine_mode new_vmode = vmode;
>      internal_fn partial_ifn = IFN_LAST;
> -     /* Produce 'len' and 'bias' argument.  */
> -     tree final_len = NULL_TREE;
> -     tree bias = NULL_TREE;
>      if (loop_lens)
>        {
>  opt_machine_mode new_ovmode
> 
 
-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)