public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: Richard Biener <rguenther@suse.de>
To: Juzhe-Zhong <juzhe.zhong@rivai.ai>
Cc: gcc-patches@gcc.gnu.org, richard.sandiford@arm.com
Subject: Re: [PATCH] gimple_fold: Support COND_LEN_FNMA/COND_LEN_FMS/COND_LEN_FNMS gimple fold
Date: Fri, 18 Aug 2023 09:26:45 +0000 (UTC)	[thread overview]
Message-ID: <nycvar.YFH.7.77.849.2308180925050.12935@jbgna.fhfr.qr> (raw)
In-Reply-To: <20230816131205.233568-1-juzhe.zhong@rivai.ai>

On Wed, 16 Aug 2023, Juzhe-Zhong wrote:

> Hi, Richard and Richi.
> 
> Currently, GCC support COND_LEN_FMA for floating-point **NO** -ffast-math.
> It's supported in tree-ssa-math-opts.cc. However, GCC failed to support COND_LEN_FNMA/COND_LEN_FMS/COND_LEN_FNMS.
> 
> Consider this following case:
> #define TEST_TYPE(TYPE)                                                        \
>   __attribute__ ((noipa)) void ternop_##TYPE (TYPE *__restrict dst,            \
> 					      TYPE *__restrict a,              \
> 					      TYPE *__restrict b, int n)       \
>   {                                                                            \
>     for (int i = 0; i < n; i++)                                                \
>       dst[i] -= a[i] * b[i];                                           \
>   }
> 
> #define TEST_ALL()                                                             \
>   TEST_TYPE (float)                                                            \
> 
> TEST_ALL ()
> 
> Gimple IR for RVV:
> 
> ...
> _39 = -vect__8.14_26;
> vect__10.16_21 = .COND_LEN_FMA ({ -1, ... }, vect__6.11_30, _39, vect__4.8_34, vect__4.8_34, _46, 0);
> ...
> 
> This is because this following piece of codes in tree-ssa-math-opts.cc:
> 
>       if (len)
> 	fma_stmt
> 	  = gimple_build_call_internal (IFN_COND_LEN_FMA, 7, cond, mulop1, op2,
> 					addop, else_value, len, bias);
>       else if (cond)
> 	fma_stmt = gimple_build_call_internal (IFN_COND_FMA, 5, cond, mulop1,
> 					       op2, addop, else_value);
>       else
> 	fma_stmt = gimple_build_call_internal (IFN_FMA, 3, mulop1, op2, addop);
>       gimple_set_lhs (fma_stmt, gimple_get_lhs (use_stmt));
>       gimple_call_set_nothrow (fma_stmt, !stmt_can_throw_internal (cfun,
> 								   use_stmt));
>       gsi_replace (&gsi, fma_stmt, true);
>       /* Follow all SSA edges so that we generate FMS, FNMA and FNMS
> 	 regardless of where the negation occurs.  */
>       gimple *orig_stmt = gsi_stmt (gsi);
>       if (fold_stmt (&gsi, follow_all_ssa_edges))
> 	{
> 	  if (maybe_clean_or_replace_eh_stmt (orig_stmt, gsi_stmt (gsi)))
> 	    gcc_unreachable ();
> 	  update_stmt (gsi_stmt (gsi));
> 	}
> 
> 'fold_stmt' failed to fold NEGATE_EXPR + COND_LEN_FMA ====> COND_LEN_FNMA.
> 
> This patch support STMT fold into:
> 
> vect__10.16_21 = .COND_LEN_FNMA ({ -1, ... }, vect__8.14_26, vect__6.11_30, vect__4.8_34, { 0.0, ... }, _46, 0);
> 
> Note that COND_LEN_FNMA has 7 arguments and COND_LEN_ADD has 6 arguments.
> 
> Extend maximum num ops:
> -  static const unsigned int MAX_NUM_OPS = 5;
> +  static const unsigned int MAX_NUM_OPS = 7;
> 
> Bootstrap and Regtest on X86 passed.
> 
> Fully tested COND_LEN_FNMA/COND_LEN_FMS/COND_LEN_FNMS on RISC-V backend.
> 
> Testing on aarch64 is on progress.
> 
> gcc/ChangeLog:
> 
>         * genmatch.cc (decision_tree::gen): Support COND_LEN_FNMA/COND_LEN_FMS/COND_LEN_FNMS gimple fold.
>         * gimple-match-exports.cc (gimple_simplify): Ditto.
>         (gimple_resimplify6): New function.
>         (gimple_resimplify7): New function.
>         (gimple_match_op::resimplify): Support COND_LEN_FNMA/COND_LEN_FMS/COND_LEN_FNMS gimple fold.
>         (convert_conditional_op): Ditto.
>         (build_call_internal): Ditto.
>         (try_conditional_simplification): Ditto.
>         (gimple_extract): Ditto.
>         * gimple-match.h (gimple_match_cond::gimple_match_cond): Ditto.
>         * internal-fn.cc (CASE): Ditto.
> 
> ---
>  gcc/genmatch.cc             |   2 +-
>  gcc/gimple-match-exports.cc | 124 ++++++++++++++++++++++++++++++++++--
>  gcc/gimple-match.h          |  19 +++++-
>  gcc/internal-fn.cc          |  11 ++--
>  4 files changed, 144 insertions(+), 12 deletions(-)
> 
> diff --git a/gcc/genmatch.cc b/gcc/genmatch.cc
> index f46d2e1520d..a1925a747a7 100644
> --- a/gcc/genmatch.cc
> +++ b/gcc/genmatch.cc
> @@ -4052,7 +4052,7 @@ decision_tree::gen (vec <FILE *> &files, bool gimple)
>      }
>    fprintf (stderr, "removed %u duplicate tails\n", rcnt);
>  
> -  for (unsigned n = 1; n <= 5; ++n)
> +  for (unsigned n = 1; n <= 7; ++n)
>      {
>        bool has_kids_p = false;
>  
> diff --git a/gcc/gimple-match-exports.cc b/gcc/gimple-match-exports.cc
> index 7aeb4ddb152..895950309b7 100644
> --- a/gcc/gimple-match-exports.cc
> +++ b/gcc/gimple-match-exports.cc
> @@ -60,6 +60,12 @@ extern bool gimple_simplify (gimple_match_op *, gimple_seq *, tree (*)(tree),
>  			     code_helper, tree, tree, tree, tree, tree);
>  extern bool gimple_simplify (gimple_match_op *, gimple_seq *, tree (*)(tree),
>  			     code_helper, tree, tree, tree, tree, tree, tree);
> +extern bool gimple_simplify (gimple_match_op *, gimple_seq *, tree (*)(tree),
> +			     code_helper, tree, tree, tree, tree, tree, tree,
> +			     tree);
> +extern bool gimple_simplify (gimple_match_op *, gimple_seq *, tree (*)(tree),
> +			     code_helper, tree, tree, tree, tree, tree, tree,
> +			     tree, tree);
>  
>  /* Functions that are needed by gimple-match but that are exported and used in
>     other places in the compiler.  */
> @@ -89,6 +95,8 @@ static bool gimple_resimplify2 (gimple_seq *, gimple_match_op *, tree (*)(tree))
>  static bool gimple_resimplify3 (gimple_seq *, gimple_match_op *, tree (*)(tree));
>  static bool gimple_resimplify4 (gimple_seq *, gimple_match_op *, tree (*)(tree));
>  static bool gimple_resimplify5 (gimple_seq *, gimple_match_op *, tree (*)(tree));
> +static bool gimple_resimplify6 (gimple_seq *, gimple_match_op *, tree (*)(tree));
> +static bool gimple_resimplify7 (gimple_seq *, gimple_match_op *, tree (*)(tree));
>  
>  /* Match and simplify the toplevel valueized operation THIS.
>     Replaces THIS with a simplified and/or canonicalized result and
> @@ -109,6 +117,10 @@ gimple_match_op::resimplify (gimple_seq *seq, tree (*valueize)(tree))
>        return gimple_resimplify4 (seq, this, valueize);
>      case 5:
>        return gimple_resimplify5 (seq, this, valueize);
> +    case 6:
> +      return gimple_resimplify6 (seq, this, valueize);
> +    case 7:
> +      return gimple_resimplify7 (seq, this, valueize);
>      default:
>        gcc_unreachable ();
>      }
> @@ -146,7 +158,14 @@ convert_conditional_op (gimple_match_op *orig_op,
>    if (ifn == IFN_LAST)
>      return false;
>    unsigned int num_ops = orig_op->num_ops;
> -  new_op->set_op (as_combined_fn (ifn), orig_op->type, num_ops + 2);
> +  unsigned int num_cond_ops = 2;
> +  if (orig_op->cond.len)
> +    {
> +      /* Convert COND_FNMA to COND_LEN_FNMA if it has LEN and BIAS.  */
> +      ifn = get_len_internal_fn (ifn);
> +      num_cond_ops = 4;
> +    }

This should definitely be placed before the ifn == IFN_LAST check,
and possibly not doing the get_conditional_internal_fn thing at all.

The genmatch changes look OK to me, I hope Richard will look at the
conditional transforms things since he invented those.

> +  new_op->set_op (as_combined_fn (ifn), orig_op->type, num_ops + num_cond_ops);
>    new_op->ops[0] = orig_op->cond.cond;
>    for (unsigned int i = 0; i < num_ops; ++i)
>      new_op->ops[i + 1] = orig_op->ops[i];
> @@ -155,6 +174,11 @@ convert_conditional_op (gimple_match_op *orig_op,
>      else_value = targetm.preferred_else_value (ifn, orig_op->type,
>  					       num_ops, orig_op->ops);
>    new_op->ops[num_ops + 1] = else_value;
> +  if (orig_op->cond.len)
> +    {
> +      new_op->ops[num_ops + 2] = orig_op->cond.len;
> +      new_op->ops[num_ops + 3] = orig_op->cond.bias;
> +    }
>    return true;
>  }
>  /* Helper for gimple_simplify valueizing OP using VALUEIZE and setting
> @@ -219,7 +243,9 @@ build_call_internal (internal_fn fn, gimple_match_op *res_op)
>  				     res_op->op_or_null (1),
>  				     res_op->op_or_null (2),
>  				     res_op->op_or_null (3),
> -				     res_op->op_or_null (4));
> +				     res_op->op_or_null (4),
> +				     res_op->op_or_null (5),
> +				     res_op->op_or_null (6));
>  }
>  
>  /* RES_OP is the result of a simplification.  If it is conditional,
> @@ -319,6 +345,7 @@ try_conditional_simplification (internal_fn ifn, gimple_match_op *res_op,
>  {
>    code_helper op;
>    tree_code code = conditional_internal_fn_code (ifn);
> +  int len_index = internal_fn_len_index (ifn);
>    if (code != ERROR_MARK)
>      op = code;
>    else
> @@ -330,12 +357,20 @@ try_conditional_simplification (internal_fn ifn, gimple_match_op *res_op,
>      }
>  
>    unsigned int num_ops = res_op->num_ops;
> +  /* num_cond_ops = 2 for COND_ADD (MASK and ELSE)
> +     wheras num_cond_ops = 4 for COND_LEN_ADD (MASK, ELSE, LEN and BIAS).  */
> +  unsigned int num_cond_ops = len_index < 0 ? 2 : 4;
> +  tree else_value
> +    = len_index < 0 ? res_op->ops[num_ops - 1] : res_op->ops[num_ops - 3];
> +  tree len = len_index < 0 ? NULL_TREE : res_op->ops[num_ops - 2];
> +  tree bias = len_index < 0 ? NULL_TREE : res_op->ops[num_ops - 1];
> +  unsigned int last_op_index = len_index < 0 ? num_ops - 1 : num_ops - 3;
>    gimple_match_op cond_op (gimple_match_cond (res_op->ops[0],
> -					      res_op->ops[num_ops - 1]),
> -			   op, res_op->type, num_ops - 2);
> +					      else_value, len, bias),
> +			   op, res_op->type, num_ops - num_cond_ops);
>  
>    memcpy (cond_op.ops, res_op->ops + 1, (num_ops - 1) * sizeof *cond_op.ops);
> -  switch (num_ops - 2)
> +  switch (num_ops - num_cond_ops)
>      {
>      case 1:
>        if (!gimple_resimplify1 (seq, &cond_op, valueize))
> @@ -717,7 +752,7 @@ gimple_extract (gimple *stmt, gimple_match_op *res_op,
>        /* ???  This way we can't simplify calls with side-effects.  */
>        if (gimple_call_lhs (stmt) != NULL_TREE
>  	  && gimple_call_num_args (stmt) >= 1
> -	  && gimple_call_num_args (stmt) <= 5)
> +	  && gimple_call_num_args (stmt) <= 7)
>  	{
>  	  combined_fn cfn;
>  	  if (gimple_call_internal_p (stmt))
> @@ -1145,6 +1180,83 @@ gimple_resimplify5 (gimple_seq *seq, gimple_match_op *res_op,
>    return canonicalized;
>  }
>  
> +/* Helper that matches and simplifies the toplevel result from
> +   a gimple_simplify run (where we don't want to build
> +   a stmt in case it's used in in-place folding).  Replaces
> +   RES_OP with a simplified and/or canonicalized result and
> +   returns whether any change was made.  */
> +
> +static bool
> +gimple_resimplify6 (gimple_seq *seq, gimple_match_op *res_op,
> +		    tree (*valueize)(tree))
> +{
> +  /* No constant folding is defined for five-operand functions.  */
> +
> +  /* Canonicalize operand order.  */
> +  bool canonicalized = false;
> +  int argno = first_commutative_argument (res_op->code, res_op->type);
> +  if (argno >= 0
> +      && tree_swap_operands_p (res_op->ops[argno], res_op->ops[argno + 1]))
> +    {
> +      std::swap (res_op->ops[argno], res_op->ops[argno + 1]);
> +      canonicalized = true;
> +    }
> +
> +  gimple_match_op res_op2 (*res_op);
> +  if (gimple_simplify (&res_op2, seq, valueize,
> +		       res_op->code, res_op->type,
> +		       res_op->ops[0], res_op->ops[1], res_op->ops[2],
> +		       res_op->ops[3], res_op->ops[4], res_op->ops[5]))
> +    {
> +      *res_op = res_op2;
> +      return true;
> +    }
> +
> +  if (maybe_resimplify_conditional_op (seq, res_op, valueize))
> +    return true;
> +
> +  return canonicalized;
> +}
> +
> +/* Helper that matches and simplifies the toplevel result from
> +   a gimple_simplify run (where we don't want to build
> +   a stmt in case it's used in in-place folding).  Replaces
> +   RES_OP with a simplified and/or canonicalized result and
> +   returns whether any change was made.  */
> +
> +static bool
> +gimple_resimplify7 (gimple_seq *seq, gimple_match_op *res_op,
> +		    tree (*valueize)(tree))
> +{
> +  /* No constant folding is defined for five-operand functions.  */
> +
> +  /* Canonicalize operand order.  */
> +  bool canonicalized = false;
> +  int argno = first_commutative_argument (res_op->code, res_op->type);
> +  if (argno >= 0
> +      && tree_swap_operands_p (res_op->ops[argno], res_op->ops[argno + 1]))
> +    {
> +      std::swap (res_op->ops[argno], res_op->ops[argno + 1]);
> +      canonicalized = true;
> +    }
> +
> +  gimple_match_op res_op2 (*res_op);
> +  if (gimple_simplify (&res_op2, seq, valueize,
> +		       res_op->code, res_op->type,
> +		       res_op->ops[0], res_op->ops[1], res_op->ops[2],
> +		       res_op->ops[3], res_op->ops[4], res_op->ops[5],
> +		       res_op->ops[6]))
> +    {
> +      *res_op = res_op2;
> +      return true;
> +    }
> +
> +  if (maybe_resimplify_conditional_op (seq, res_op, valueize))
> +    return true;
> +
> +  return canonicalized;
> +}
> +
>  /* Return a canonical form for CODE when operating on TYPE.  The idea
>     is to remove redundant ways of representing the same operation so
>     that code_helpers can be hashed and compared for equality.
> diff --git a/gcc/gimple-match.h b/gcc/gimple-match.h
> index b20585dca4b..c16cfede826 100644
> --- a/gcc/gimple-match.h
> +++ b/gcc/gimple-match.h
> @@ -34,6 +34,7 @@ public:
>    /* Build an unconditional op.  */
>    gimple_match_cond (uncond) : cond (NULL_TREE), else_value (NULL_TREE) {}
>    gimple_match_cond (tree, tree);
> +  gimple_match_cond (tree, tree, tree, tree);
>  
>    gimple_match_cond any_else () const;
>  
> @@ -44,6 +45,16 @@ public:
>    /* The value to use when the condition is false.  This is NULL_TREE if
>       the operation is unconditional or if the value doesn't matter.  */
>    tree else_value;
> +
> +  /* The length value for conditional len operation that operation occurs
> +     when element index < len + bias.  This is NULL_TREE if the operation
> +     is non-LEN operation.  */
> +  tree len;
> +
> +  /* The bias value for conditional len operation that operation occurs
> +     when element index < len + bias.  This is NULL_TREE if the operation
> +     is non-LEN operation.  */
> +  tree bias;
>  };
>  
>  inline
> @@ -52,6 +63,12 @@ gimple_match_cond::gimple_match_cond (tree cond_in, tree else_value_in)
>  {
>  }
>  
> +inline
> +gimple_match_cond::gimple_match_cond (tree cond_in, tree else_value_in,
> +				      tree len_in, tree bias_in)
> +  : cond (cond_in), else_value (else_value_in), len (len_in), bias (bias_in)
> +{}
> +
>  /* Return a gimple_match_cond with the same condition but with an
>     arbitrary ELSE_VALUE.  */
>  
> @@ -93,7 +110,7 @@ public:
>    bool resimplify (gimple_seq *, tree (*)(tree));
>  
>    /* The maximum value of NUM_OPS.  */
> -  static const unsigned int MAX_NUM_OPS = 5;
> +  static const unsigned int MAX_NUM_OPS = 7;
>  
>    /* The conditions under which the operation is performed, and the value to
>       use as a fallback.  */
> diff --git a/gcc/internal-fn.cc b/gcc/internal-fn.cc
> index cc1ede58799..be9b93e9f71 100644
> --- a/gcc/internal-fn.cc
> +++ b/gcc/internal-fn.cc
> @@ -4477,11 +4477,14 @@ get_unconditional_internal_fn (internal_fn ifn)
>  {
>    switch (ifn)
>      {
> -#define CASE(NAME) case IFN_COND_##NAME: return IFN_##NAME;
> -      FOR_EACH_COND_FN_PAIR(CASE)
> +#define CASE(NAME)                                                             \
> +  case IFN_COND_##NAME:                                                        \
> +  case IFN_COND_LEN_##NAME:                                                    \
> +    return IFN_##NAME;
> +      FOR_EACH_COND_FN_PAIR (CASE)
>  #undef CASE
> -    default:
> -      return IFN_LAST;
> +      default:
> +	return IFN_LAST;
>      }
>  }
>  
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)

  reply	other threads:[~2023-08-18  9:26 UTC|newest]

Thread overview: 6+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2023-08-16 13:12 Juzhe-Zhong
2023-08-18  9:26 ` Richard Biener [this message]
2023-08-18 10:06   ` Juzhe Zhong
2023-08-21 14:18   ` Richard Sandiford
2023-08-21 14:16 ` Richard Sandiford
2023-08-22  3:00   ` juzhe.zhong

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=nycvar.YFH.7.77.849.2308180925050.12935@jbgna.fhfr.qr \
    --to=rguenther@suse.de \
    --cc=gcc-patches@gcc.gnu.org \
    --cc=juzhe.zhong@rivai.ai \
    --cc=richard.sandiford@arm.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).