vect: Fix wrong check in vect_recog_mulhs_pattern [PR101596]

diff --git a/gcc/testsuite/gcc.target/powerpc/pr101596-1.c b/gcc/testsuite/gcc.target/powerpc/pr101596-1.c
new file mode 100644
index 00000000000..ed09e5f96ed
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/pr101596-1.c
@@ -0,0 +1,30 @@ 
+/* { dg-require-effective-target power10_ok } */
+/* { dg-options "-mdejagnu-cpu=power10 -O2 -ftree-vectorize -fno-vect-cost-model -fdump-tree-vect-details" } */
+
+/* Check vect_recog_mulhs_pattern can't be detected with shift count 48.  */
+
+#define N 128
+
+typedef signed long long sLL;
+typedef unsigned long long uLL;
+
+signed int si_a[N], si_b[N];
+unsigned int ui_a[N], ui_b[N];
+signed short sh_c[N];
+unsigned short uh_c[N];
+
+void
+test1 ()
+{
+  for (int i = 0; i < N; i++)
+    sh_c[i] = ((sLL) si_a[i] * (sLL) si_b[i]) >> 48;
+}
+
+void
+test2 ()
+{
+  for (int i = 0; i < N; i++)
+    uh_c[i] = ((uLL) ui_a[i] * (uLL) ui_b[i]) >> 48;
+}
+
+/* { dg-final { scan-tree-dump-not "vect_recog_mulhs_pattern: detected" "vect" } } */
diff --git a/gcc/testsuite/gcc.target/powerpc/pr101596-2.c b/gcc/testsuite/gcc.target/powerpc/pr101596-2.c
new file mode 100644
index 00000000000..5caa7ce4189
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/pr101596-2.c
@@ -0,0 +1,30 @@ 
+/* { dg-require-effective-target power10_ok } */
+/* { dg-options "-mdejagnu-cpu=power10 -O2 -ftree-vectorize -fno-vect-cost-model -fdump-tree-vect-details" } */
+
+/* Check vect_recog_mulhs_pattern can be detected with shift count 32.  */
+
+#define N 128
+
+typedef signed long long sLL;
+typedef unsigned long long uLL;
+
+signed int si_a[N], si_b[N];
+unsigned int ui_a[N], ui_b[N];
+signed short sh_c[N];
+unsigned short uh_c[N];
+
+void
+test1 ()
+{
+  for (int i = 0; i < N; i++)
+    sh_c[i] = ((sLL) si_a[i] * (sLL) si_b[i]) >> 32;
+}
+
+void
+test2 ()
+{
+  for (int i = 0; i < N; i++)
+    uh_c[i] = ((uLL) ui_a[i] * (uLL) ui_b[i]) >> 32;
+}
+
+/* { dg-final { scan-tree-dump-times "vect_recog_mulhs_pattern: detected" 2 "vect" } } */
diff --git a/gcc/testsuite/gcc.target/powerpc/pr101596-3.c b/gcc/testsuite/gcc.target/powerpc/pr101596-3.c
new file mode 100644
index 00000000000..4b92661412d
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/pr101596-3.c
@@ -0,0 +1,58 @@ 
+/* { dg-do run } */
+/* { dg-require-effective-target power10_hw } */
+/* { dg-options "-mdejagnu-cpu=power10 -O2 -ftree-vectorize -fno-vect-cost-model" } */
+
+/* Verify the execution goes well with shift count either 32 or 48.  */
+
+#define N 128
+
+typedef signed int si;
+typedef signed short sh;
+typedef signed long long sll;
+typedef unsigned int ui;
+typedef unsigned short uh;
+typedef unsigned long long ull;
+
+si si_a[N], si_b[N];
+ui ui_a[N], ui_b[N];
+sh sh_c[N];
+uh uh_c[N];
+
+#define TEST(NTYPE, TYPE, WTYPE, CNT)                                          \
+  void __attribute__ ((noipa)) test_##TYPE##CNT ()                             \
+  {                                                                            \
+    for (int i = 0; i < N; i++)                                                \
+      NTYPE##_c[i] = ((WTYPE) TYPE##_a[i] * (WTYPE) TYPE##_b[i]) >> CNT;       \
+  }                                                                            \
+                                                                               \
+  void __attribute__ ((noipa, optimize ("O1"))) check_##TYPE##CNT ()           \
+  {                                                                            \
+    test_##TYPE##CNT ();                                                       \
+    for (int i = 0; i < N; i++)                                                \
+      {                                                                        \
+	NTYPE exp = ((WTYPE) TYPE##_a[i] * (WTYPE) TYPE##_b[i]) >> CNT;        \
+	if (NTYPE##_c[i] != exp)                                               \
+	  __builtin_abort ();                                                  \
+      }                                                                        \
+  }
+
+TEST (sh, si, sll, 32)
+TEST (sh, si, sll, 48)
+TEST (uh, ui, ull, 32)
+TEST (uh, ui, ull, 48)
+
+int
+main ()
+{
+
+  for (int i = 0; i < N; i++)
+    {
+      ui_a[i] = si_a[i] = 0x12345678ULL + 0x1000ULL * (i * 3 - 1);
+      ui_b[i] = si_b[i] = 0x87654321ULL - 0x500000ULL * (i * 5 + 1);
+    }
+
+  check_si32 ();
+  check_si48 ();
+  check_ui32 ();
+  check_ui48 ();
+}
diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c
index 70bb751865b..743fd3f5414 100644
--- a/gcc/tree-vect-patterns.c
+++ b/gcc/tree-vect-patterns.c
@@ -1986,7 +1986,7 @@  vect_recog_mulhs_pattern (vec_info *vinfo,
 
   stmt_vec_info mulh_stmt_info;
   tree scale_term;
-  internal_fn ifn;
+  bool rounding_p = false;
 
   /* Check for the presence of the rounding term.  */
   if (gimple_assign_rhs_code (rshift_input_stmt) == PLUS_EXPR)
@@ -2035,37 +2035,18 @@  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;
-
-      ifn = IFN_MULHRS;
+      rounding_p = true;
     }
   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;
-
-      /* 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_MULH;
-      else
-	return NULL;
     }
 
+  /* Check that the scaling factor is constant.  */
+  if (TREE_CODE (scale_term) != INTEGER_CST)
+    return NULL;
+
   /* Check whether the scaling input term can be seen as two widened
      inputs multiplied together.  */
   vect_unpromoted_value unprom_mult[2];
@@ -2076,13 +2057,41 @@  vect_recog_mulhs_pattern (vec_info *vinfo,
   if (nops != 2)
     return NULL;
 
-  vect_pattern_detected ("vect_recog_mulhs_pattern", last_stmt);
-
   /* Adjust output precision.  */
   if (TYPE_PRECISION (new_type) < target_precision)
     new_type = build_nonstandard_integer_type
       (target_precision, TYPE_UNSIGNED (new_type));
 
+  unsigned mult_precision = TYPE_PRECISION (new_type);
+  internal_fn ifn;
+  /* Check that the scaling factor is expected.  Instead of
+     target_precision, we should use the one that we actually
+     use for internal function.  */
+  if (rounding_p)
+    {
+      /* Check pattern 2).  */
+      if (wi::to_widest (scale_term) + mult_precision + 2
+	  != TYPE_PRECISION (lhs_type))
+	return NULL;
+
+      ifn = IFN_MULHRS;
+    }
+  else
+    {
+      /* Check for pattern 1).  */
+      if (wi::to_widest (scale_term) + mult_precision + 1
+	  == TYPE_PRECISION (lhs_type))
+	ifn = IFN_MULHS;
+      /* Check for pattern 3).  */
+      else if (wi::to_widest (scale_term) + mult_precision
+	       == TYPE_PRECISION (lhs_type))
+	ifn = IFN_MULH;
+      else
+	return NULL;
+    }
+
+  vect_pattern_detected ("vect_recog_mulhs_pattern", last_stmt);
+
   /* Check for target support.  */
   tree new_vectype = get_vectype_for_scalar_type (vinfo, new_type);
   if (!new_vectype