vectorizer: Fix up vectorization using WIDEN_MINUS_EXPR [PR102124]

The following testcase is miscompiled on aarch64-linux at -O3 since the
introduction of WIDEN_MINUS_EXPR.
The problem is if the inner type (half_type) is unsigned and the result
type in which the subtraction is performed (type) has precision more than
twice as larger as the inner type's precision.
For other widening operations like WIDEN_{PLUS,MULT}_EXPR, if half_type
is unsigned, the addition/multiplication result in itype is also unsigned
and needs to be zero-extended to type.
But subtraction is special, even when half_type is unsigned, the subtraction
behaves as signed (also regardless of whether the result type is signed or
unsigned), 0xfeU - 0xffU is -1 or 0xffffffffU, not 0x0000ffff.

I think it is better not to use mixed signedness of types in
WIDEN_MINUS_EXPR (have unsigned vector of operands and signed result
vector), so this patch instead adds another cast to make sure we always
sign-extend the result from itype to type if type is wider than itype.

2021-09-01  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/102124
	* tree-vect-patterns.c (vect_recog_widen_op_pattern): For ORIG_CODE
	MINUS_EXPR, if itype is unsigned with smaller precision than type,
	add an extra cast to signed variant of itype to ensure sign-extension.

	* gcc.dg/torture/pr102124.c: New test.

1 files changed, 25 insertions, 1 deletions

diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c
index 8997340..e6c5bcd 100644
--- a/gcc/tree-vect-patterns.c
+++ b/gcc/tree-vect-patterns.c
@@ -1268,11 +1268,31 @@ vect_recog_widen_op_pattern (vec_info *vinfo,
   /* Check target support  */
   tree vectype = get_vectype_for_scalar_type (vinfo, half_type);
   tree vecitype = get_vectype_for_scalar_type (vinfo, itype);
+  tree ctype = itype;
+  tree vecctype = vecitype;
+  if (orig_code == MINUS_EXPR
+      && TYPE_UNSIGNED (itype)
+      && TYPE_PRECISION (type) > TYPE_PRECISION (itype))
+    {
+      /* Subtraction is special, even if half_type is unsigned and no matter
+	 whether type is signed or unsigned, if type is wider than itype,
+	 we need to sign-extend from the widening operation result to the
+	 result type.
+	 Consider half_type unsigned char, operand 1 0xfe, operand 2 0xff,
+	 itype unsigned short and type either int or unsigned int.
+	 Widened (unsigned short) 0xfe - (unsigned short) 0xff is
+	 (unsigned short) 0xffff, but for type int we want the result -1
+	 and for type unsigned int 0xffffffff rather than 0xffff.  */
+      ctype = build_nonstandard_integer_type (TYPE_PRECISION (itype), 0);
+      vecctype = get_vectype_for_scalar_type (vinfo, ctype);
+    }
+
   enum tree_code dummy_code;
   int dummy_int;
   auto_vec<tree> dummy_vec;
   if (!vectype
       || !vecitype
+      || !vecctype
       || !supportable_widening_operation (vinfo, wide_code, last_stmt_info,
 					  vecitype, vectype,
 					  &dummy_code, &dummy_code,
@@ -1291,8 +1311,12 @@ vect_recog_widen_op_pattern (vec_info *vinfo,
   gimple *pattern_stmt = gimple_build_assign (var, wide_code,
 					      oprnd[0], oprnd[1]);
 
+  if (vecctype != vecitype)
+    pattern_stmt = vect_convert_output (vinfo, last_stmt_info, ctype,
+					pattern_stmt, vecitype);
+
   return vect_convert_output (vinfo, last_stmt_info,
-			      type, pattern_stmt, vecitype);
+			      type, pattern_stmt, vecctype);
 }
 
 /* Try to detect multiplication on widened inputs, converting MULT_EXPR