Re: Re: [PATCH] VECT: Support CALL vectorization for COND_LEN_*

[钟居哲 <juzhe.zhong@rivai.ai>]

[-- Attachment #1: Type: text/plain, Size: 12384 bytes --]

Hi, Richi. Thank you so much for review.

>> This function doesn't seem to care about conditional vectorization
>> support, so why are you changing it?

I debug and analyze the code here:

Breakpoint 1, vectorizable_call (vinfo=0x3d358d0, stmt_info=0x3dcc820, gsi=0x0, vec_stmt=0x0, slp_node=0x0, cost_vec=0x7fffffffc668) at ../../../riscv-gcc/gcc/tree-vect-stmts.cc:3485
3485        ifn = vectorizable_internal_function (cfn, callee, vectype_out,
(gdb) p cfn
$1 = CFN_COND_ADD
(gdb) call print_gimple_stmt(stdout,stmt,0,0)
_9 = .COND_ADD (_27, _6, _8, _6);

When target like RISC-V didnt' support COND_* (we only support COND_LEN_*, no support COND_*),
ifn will be IFN_LAST.
Also, the following code doesn't modify ifn until
"internal_fn cond_fn = get_conditional_internal_fn (ifn);"
When ifn == IFN_LAST, cond_fn will IFN_LAST too.

So, I extend "vectorizable_internal_function" to return COND_LEN_*.

Am I going into wrong direction on debug && analysis.

>> what about reduc_indx?  Why check len_indes and not len_loop_p?

>> Note traditionally non-LEN fns would return -1 here, so ...

>> instead there's code here that checks for the .COND_xxx case so you should
>> ament it for the len case as well?

>> You are not actually using len_index to place the arguments.  See
>> how mask_opno is used.

I understand this patch codes look quite strange not natural.
Give me some time to share the story why I am doing this:

1. At the beginning, I want to write the codes naturally and very similiar with the current codes.
    So it's natural to write the codes like this:
     
     if (mask_opno >= 0 && len_loop_p) {
         tree len = vect_get_loop_len
         tree bias = ....
      }

    However, we have the following case:
    void foo1 (float * __restrict a, float * __restrict b, int * __restrict cond, int n)
    {
       for (int i = 0; i < 8; i++)
        if (cond[i])
           a[i] = b[i] + a[i];
     }     
     Note here "i<8" means niters = 8 and compile with known VF (fixed-length vector) to vectorize the codes.
     Then len_loop_p will be false.

     If I don't handle case, then the code will generate the IR:
     COND_LEN_ADD (maks, op1, op2, op3).... 
     Missing LEN && BIAS argument which cause ICE.

2. So I use "int len_index = internal_fn_len_index (ifn);",
    ifn is supported as COND_LEN_*, the len_index >= 0, otherwise, it will be -1.
     So I prepare LEN && BIAS arguments according to len_index.
     
     So I change the code into:
      if (len_index >= 0) {
         tree len = vect_get_loop_len
         tree bias = ....
      }
       
3. Overall, I want to normalize COND_* into COND_LEN_*, when loop_len_p is false, then I create a dummy LEN for it.
    It make the codes strange. 
    
4. I have another approach/prepared patch which makes codes more natural:
    Enable COND_xxxx in RISC-V port, then keep "tree-vect-ifcvt.cc" unchange,
    and "vectorizable_internal_function" unchange.
    
    And simpily write the codes like this:
     if (mask_opno >= 0 && len_loop_p) {
         tree len = vect_get_loop_len
         tree bias = ....
      }
    and when len_loop_p is true, I change to use "COND_LEN_xxx", otherwise reuse original flow.
    
    Are you suggesting me doing this? Meaning enable COND_xxx (The patterns are duplicate I think if we have COND_LEN_xxx) in RISC-V port ?

>> You are producing patches too quickly, please double-check what you did
>> before posting for review.

I am so sorry this patch codes look quite strange and confusing.
I actually consider it for some time
(I have 2 prepared patches, this one which is not natural, the other only add a very few codes and look natural and more acceptable ), 
and compare them (the other one I need to support COND_xxx in RISC-V port).

Maybe it's better enable COND_xxx in RISC-V port so that the codes in middle-end look much more reasonable ?
I think maybe next time I should ask you first about the solution before I send the patch

Looking forward your suggestions.
Thanks.


juzhe.zhong@rivai.ai
 
From: Richard Biener
Date: 2023-07-24 22:33
To: Ju-Zhe Zhong
CC: gcc-patches; richard.sandiford
Subject: Re: [PATCH] VECT: Support CALL vectorization for COND_LEN_*
On Mon, 24 Jul 2023, juzhe.zhong@rivai.ai wrote:
 
> From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
> 
> Hi, Richard and Richi.
> 
> This patch supports CALL vectorization by COND_LEN_*.
> 
> Consider this following case:
> void foo (float * __restrict a, float * __restrict b, int * __restrict cond, int n)
> {
>   for (int i = 0; i < n; i++)
>     if (cond[i])
>       a[i] = b[i] + a[i];
> }
> 
> 
> Output of RISC-V (32-bits) gcc (trunk) (Compiler #3)
> <source>:5:21: missed: couldn't vectorize loop
> <source>:5:21: missed: not vectorized: control flow in loop.
> 
> ARM SVE:
> 
> ...
> mask__27.10_51 = vect__4.9_49 != { 0, ... };
> ...
> vec_mask_and_55 = loop_mask_49 & mask__27.10_51;
> ...
> vect__9.17_62 = .COND_ADD (vec_mask_and_55, vect__6.13_56, vect__8.16_60, vect__6.13_56);
> 
> For RVV, we want IR as follows:
> 
> ...
> _68 = .SELECT_VL (ivtmp_66, POLY_INT_CST [4, 4]);
> ...
> mask__27.10_51 = vect__4.9_49 != { 0, ... };
> ...
> vect__9.17_60 = .COND_LEN_ADD (mask__27.10_51, vect__6.13_55, vect__8.16_59, vect__6.13_55, _68, 0);
> ...
> 
> Both len and mask of COND_LEN_ADD are real not dummy.
> 
> gcc/ChangeLog:
> 
>         * tree-if-conv.cc (ifcvt_can_predicate): Enable ifcvt for COND_LEN_*.
>         * tree-vect-stmts.cc (vectorizable_internal_function): Apply COND_LEN_*.
>         (vectorizable_call): Ditto.
> 
> ---
>  gcc/tree-if-conv.cc    |  7 +++-
>  gcc/tree-vect-stmts.cc | 84 +++++++++++++++++++++++++++++++++++++++---
>  2 files changed, 83 insertions(+), 8 deletions(-)
> 
> diff --git a/gcc/tree-if-conv.cc b/gcc/tree-if-conv.cc
> index 799f071965e..2e17658b9ed 100644
> --- a/gcc/tree-if-conv.cc
> +++ b/gcc/tree-if-conv.cc
> @@ -1010,8 +1010,11 @@ ifcvt_can_predicate (gimple *stmt)
>    if (!types_compatible_p (lhs_type, rhs_type))
>      return false;
>    internal_fn cond_fn = get_conditional_internal_fn (code);
> -  return (cond_fn != IFN_LAST
> -   && vectorized_internal_fn_supported_p (cond_fn, lhs_type));
> +  internal_fn cond_len_fn = get_conditional_len_internal_fn (code);
> +  return ((cond_fn != IFN_LAST
> +    && vectorized_internal_fn_supported_p (cond_fn, lhs_type))
> +   || (cond_len_fn != IFN_LAST
> +       && vectorized_internal_fn_supported_p (cond_len_fn, lhs_type)));
>  }
>  
>  /* Return true when STMT is if-convertible.
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index ed28fbdced3..15401ddb682 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -1543,10 +1543,27 @@ vectorizable_internal_function (combined_fn cfn, tree fndecl,
>  tree vectype_out, tree vectype_in)
>  {
>    internal_fn ifn;
> +  internal_fn len_ifn = IFN_LAST;
>    if (internal_fn_p (cfn))
> -    ifn = as_internal_fn (cfn);
> +    {
> +      ifn = as_internal_fn (cfn);
> +      tree_code code = conditional_internal_fn_code (ifn);
> +      len_ifn = get_conditional_len_internal_fn (code);
> +    }
>    else
>      ifn = associated_internal_fn (fndecl);
 
This function doesn't seem to care about conditional vectorization
support, so why are you changing it?
 
> +  if (len_ifn != IFN_LAST && direct_internal_fn_p (len_ifn))
> +    {
> +      const direct_internal_fn_info &info = direct_internal_fn (len_ifn);
> +      if (info.vectorizable)
> + {
> +   tree type0 = (info.type0 < 0 ? vectype_out : vectype_in);
> +   tree type1 = (info.type1 < 0 ? vectype_out : vectype_in);
> +   if (direct_internal_fn_supported_p (len_ifn, tree_pair (type0, type1),
> +       OPTIMIZE_FOR_SPEED))
> +     return len_ifn;
> + }
> +    }
>    if (ifn != IFN_LAST && direct_internal_fn_p (ifn))
>      {
>        const direct_internal_fn_info &info = direct_internal_fn (ifn);
> @@ -3538,8 +3555,10 @@ vectorizable_call (vec_info *vinfo,
>    gcc_assert (ncopies >= 1);
>  
>    int reduc_idx = STMT_VINFO_REDUC_IDX (stmt_info);
> +  int len_index = internal_fn_len_index (ifn);
 
Note traditionally non-LEN fns would return -1 here, so ...
 
>    internal_fn cond_fn = get_conditional_internal_fn (ifn);
>    vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL);
> +  vec_loop_lens *lens = (loop_vinfo ? &LOOP_VINFO_LENS (loop_vinfo) : NULL);
>    if (!vec_stmt) /* transformation not required.  */
>      {
>        if (slp_node)
> @@ -3585,8 +3604,12 @@ vectorizable_call (vec_info *vinfo,
 
instead there's code here that checks for the .COND_xxx case so you should
ament it for the len case as well?
 
>        tree scalar_mask = NULL_TREE;
>        if (mask_opno >= 0)
>  scalar_mask = gimple_call_arg (stmt_info->stmt, mask_opno);
> -       vect_record_loop_mask (loop_vinfo, masks, nvectors,
> -      vectype_out, scalar_mask);
> +       if (len_index >= 0)
> + vect_record_loop_len (loop_vinfo, lens, nvectors, vectype_out,
> +       1);
> +       else
> + vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype_out,
> +        scalar_mask);
>      }
>  }
>        return true;
> @@ -3602,8 +3625,16 @@ vectorizable_call (vec_info *vinfo,
>    vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
>  
>    bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
> +  bool len_loop_p = loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo);
>    unsigned int vect_nargs = nargs;
> -  if (masked_loop_p && reduc_idx >= 0)
> +  if (len_index)
 
what about reduc_indx?  Why check len_indes and not len_loop_p?
 
... checking len_index is going to be true for -1, you probably would
want len_index != -1 then?
 
> +    {
> +      /* COND_ADD --> COND_LEN_ADD with 2 more args (LEN + BIAS).
> + FIXME: We don't support FMAX --> COND_LEN_FMAX yet which
> + needs 4 more args (MASK + ELSE + LEN + BIAS).  */
> +      vect_nargs += 2;
> +    }
> +  else if (masked_loop_p && reduc_idx >= 0)
>      {
>        ifn = cond_fn;
>        vect_nargs += 2;
> @@ -3670,7 +3701,29 @@ vectorizable_call (vec_info *vinfo,
>      }
>    else
>      {
> -       if (mask_opno >= 0 && masked_loop_p)
> +       if (len_index)
 
see above
 
> + {
> +   tree len, bias;
> +   if (len_loop_p)
> +     len
> +       = vect_get_loop_len (loop_vinfo, gsi, lens,
> +    ncopies, vectype_out, j, 1);
> +   else
> +     {
> +       /* Dummy LEN.  */
> +       tree iv_type
> + = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> +       len
> + = build_int_cst (iv_type, TYPE_VECTOR_SUBPARTS (
> +     vectype_out));
> +     }
> +   signed char biasval
> +     = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
> +   bias = build_int_cst (intQI_type_node, biasval);
> +   vargs[varg++] = len;
> +   vargs[varg++] = bias;
> + }
> +       else if (mask_opno >= 0 && masked_loop_p)
>  {
>    unsigned int vec_num = vec_oprnds0.length ();
>    /* Always true for SLP.  */
> @@ -3718,7 +3771,26 @@ vectorizable_call (vec_info *vinfo,
>    if (masked_loop_p && reduc_idx >= 0)
>      vargs[varg++] = vargs[reduc_idx + 1];
>  
> -   if (mask_opno >= 0 && masked_loop_p)
> +   if (len_index)
> +     {
> +       tree len, bias;
> +       if (len_loop_p)
> + len = vect_get_loop_len (loop_vinfo, gsi, lens, ncopies,
> + vectype_out, j, 1);
> +       else
> + {
> +   /* Dummy LEN.  */
> +   tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> +   len = build_int_cst (iv_type,
> +        TYPE_VECTOR_SUBPARTS (vectype_out));
> + }
> +       signed char biasval
> + = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
> +       bias = build_int_cst (intQI_type_node, biasval);
> +       vargs[varg++] = len;
> +       vargs[varg++] = bias;
 
 
You are not actually using len_index to place the arguments.  See
how mask_opno is used.
 
You are producing patches too quickly, please double-check what you did
before posting for review.
 
Thanks,
Richard.
 
> +     }
> +   else if (mask_opno >= 0 && masked_loop_p)
>      {
>        tree mask = vect_get_loop_mask (loop_vinfo, gsi, masks, ncopies,
>        vectype_out, j);
>