gimple_fold: Support COND_LEN_FNMA/COND_LEN_FMS/COND_LEN_FNMS gimple fold

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:
  __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];                                           \
  }

  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.
Tested on aarch64 Qemu.

Fully tested COND_LEN_FNMA/COND_LEN_FMS/COND_LEN_FNMS on RISC-V backend.

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.

diff --git a/gcc/genmatch.cc b/gcc/genmatch.cc
index f46d2e1520d0daee1f373af47846037297ffa97d..a1925a747a7351ea4c5168bccaa857436f93ed93 100644 (file)
--- 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 7aeb4ddb152468ddef19e6361428498371a1ebf2..b36027b0bad783e3180d0ec74deb3a07b12906fc 100644 (file)
--- 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)
+    {
+      /* Add the length and bias parameters.  */
+      ifn = get_len_internal_fn (ifn);
+      num_cond_ops = 4;
+    }
+  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 +357,19 @@ 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];
   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 +751,7 @@ 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 +1179,83 @@ 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 six-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 seven-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 b20585dca4b0204ec4b332a89c2737a82d3f7d4d..bec3ff42e3ecb197c3663898da39acec8a4dd39e 100644 (file)
--- 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 +45,13 @@ 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 and bias parameters to be applied to a vector operation,
+     so that the condition is forced to false when the element index is
+     >= LEN + BIAS.  These are NULL_TREE if the operation isn't applied
+     to vectors, or if no such length limit is in use.  */
+  tree len;
+  tree bias;
 };
 
 inline
@@ -52,6 +60,12 @@ 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 +107,7 @@ 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 8baf4d4cd5179a4cc2c81ae95b1b8c49abc4b5d5..4138cc31d7eda6f5b2550a968c2b0f15c5f9e9ca 100644 (file)
--- a/gcc/internal-fn.cc
+++ b/gcc/internal-fn.cc
@@ -4448,8 +4448,11 @@ 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;