[RFC/PATCH] vect: Recog mul_highpart pattern

["Kewen.Lin" <linkw@linux.ibm.com>]

[-- Attachment #1: Type: text/plain, Size: 1030 bytes --]

Hi,

When I added the support for Power10 newly introduced multiply
highpart instrutions, I noticed that currently vectorizer
doesn't try to vectorize multiply highpart pattern, I hope
this isn't intentional?

This patch is to extend the existing pattern mulhs handlings
to cover multiply highpart.  Another alternative seems to
recog mul_highpart operation in a general place applied for
scalar code when the target supports the optab for the scalar
operation, it's based on the assumption that one target which
supports vector version of multiply highpart should have the
scalar version.  I noticed that the function can_mult_highpart_p
can check/handle mult_highpart well even without mul_highpart
optab support, I think to recog this pattern in vectorizer
is better.  Is it on the right track?

Bootstrapped & regtested on powerpc64le-linux-gnu P9,
x86_64-redhat-linux and aarch64-linux-gnu.

BR,
Kewen
-----
gcc/ChangeLog:

	* tree-vect-patterns.c (vect_recog_mulhs_pattern): Add support to
	recog normal multiply highpart.


[-- Attachment #2: 0001-vect-Recog-mul_highpart-pattern.patch --]
[-- Type: text/plain, Size: 4279 bytes --]

---
 gcc/tree-vect-patterns.c | 67 ++++++++++++++++++++++++++++------------
 1 file changed, 48 insertions(+), 19 deletions(-)

diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c
index b2e7fc2cc7a..9253c8088e9 100644
--- a/gcc/tree-vect-patterns.c
+++ b/gcc/tree-vect-patterns.c
@@ -1896,8 +1896,15 @@ vect_recog_over_widening_pattern (vec_info *vinfo,
 
    1) Multiply high with scaling
      TYPE res = ((TYPE) a * (TYPE) b) >> c;
+     Here, c is bitsize (TYPE) / 2 - 1.
+
    2) ... or also with rounding
      TYPE res = (((TYPE) a * (TYPE) b) >> d + 1) >> 1;
+     Here, d is bitsize (TYPE) / 2 - 2.
+
+   3) Normal multiply high
+     TYPE res = ((TYPE) a * (TYPE) b) >> e;
+     Here, e is bitsize (TYPE) / 2.
 
    where only the bottom half of res is used.  */
 
@@ -1942,7 +1949,6 @@ vect_recog_mulhs_pattern (vec_info *vinfo,
   stmt_vec_info mulh_stmt_info;
   tree scale_term;
   internal_fn ifn;
-  unsigned int expect_offset;
 
   /* Check for the presence of the rounding term.  */
   if (gimple_assign_rhs_code (rshift_input_stmt) == PLUS_EXPR)
@@ -1991,25 +1997,37 @@ vect_recog_mulhs_pattern (vec_info *vinfo,
 
       /* Get the scaling term.  */
       scale_term = gimple_assign_rhs2 (plus_input_stmt);
+      /* Check that the scaling factor is correct.  */
+      if (TREE_CODE (scale_term) != INTEGER_CST)
+	return NULL;
+
+      /* Check pattern 2).  */
+      if (wi::to_widest (scale_term) + target_precision + 2
+	  != TYPE_PRECISION (lhs_type))
+	return NULL;
 
-      expect_offset = target_precision + 2;
       ifn = IFN_MULHRS;
     }
   else
     {
       mulh_stmt_info = rshift_input_stmt_info;
       scale_term = gimple_assign_rhs2 (last_stmt);
+      /* Check that the scaling factor is correct.  */
+      if (TREE_CODE (scale_term) != INTEGER_CST)
+	return NULL;
 
-      expect_offset = target_precision + 1;
-      ifn = IFN_MULHS;
+      /* Check for pattern 1).  */
+      if (wi::to_widest (scale_term) + target_precision + 1
+	  == TYPE_PRECISION (lhs_type))
+	ifn = IFN_MULHS;
+      /* Check for pattern 3).  */
+      else if (wi::to_widest (scale_term) + target_precision
+	       == TYPE_PRECISION (lhs_type))
+	ifn = IFN_LAST;
+      else
+	return NULL;
     }
 
-  /* Check that the scaling factor is correct.  */
-  if (TREE_CODE (scale_term) != INTEGER_CST
-      || wi::to_widest (scale_term) + expect_offset
-	   != TYPE_PRECISION (lhs_type))
-    return NULL;
-
   /* Check whether the scaling input term can be seen as two widened
      inputs multiplied together.  */
   vect_unpromoted_value unprom_mult[2];
@@ -2029,9 +2047,14 @@ vect_recog_mulhs_pattern (vec_info *vinfo,
 
   /* Check for target support.  */
   tree new_vectype = get_vectype_for_scalar_type (vinfo, new_type);
-  if (!new_vectype
-      || !direct_internal_fn_supported_p
-	    (ifn, new_vectype, OPTIMIZE_FOR_SPEED))
+  if (!new_vectype)
+    return NULL;
+  if (ifn != IFN_LAST
+      && !direct_internal_fn_supported_p (ifn, new_vectype, OPTIMIZE_FOR_SPEED))
+    return NULL;
+  else if (ifn == IFN_LAST
+	   && !can_mult_highpart_p (TYPE_MODE (new_vectype),
+				    TYPE_UNSIGNED (new_type)))
     return NULL;
 
   /* The IR requires a valid vector type for the cast result, even though
@@ -2040,14 +2063,20 @@ vect_recog_mulhs_pattern (vec_info *vinfo,
   if (!*type_out)
     return NULL;
 
-  /* Generate the IFN_MULHRS call.  */
+  gimple *mulhrs_stmt;
   tree new_var = vect_recog_temp_ssa_var (new_type, NULL);
   tree new_ops[2];
-  vect_convert_inputs (vinfo, last_stmt_info, 2, new_ops, new_type,
-		       unprom_mult, new_vectype);
-  gcall *mulhrs_stmt
-    = gimple_build_call_internal (ifn, 2, new_ops[0], new_ops[1]);
-  gimple_call_set_lhs (mulhrs_stmt, new_var);
+  vect_convert_inputs (vinfo, last_stmt_info, 2, new_ops, new_type, unprom_mult,
+		       new_vectype);
+  if (ifn == IFN_LAST)
+    mulhrs_stmt = gimple_build_assign (new_var, MULT_HIGHPART_EXPR, new_ops[0],
+				       new_ops[1]);
+  else
+    {
+      /* Generate the IFN_MULHRS call.  */
+      mulhrs_stmt = gimple_build_call_internal (ifn, 2, new_ops[0], new_ops[1]);
+      gimple_call_set_lhs (mulhrs_stmt, new_var);
+    }
   gimple_set_location (mulhrs_stmt, gimple_location (last_stmt));
 
   if (dump_enabled_p ())
-- 
2.17.1