On January 13, 2016 4:26:50 PM GMT+01:00, Marek Polacek <pola...@redhat.com>
wrote:
>We crash on the following testcase because the FEs create
>
> VEC_COND_EXPR < a == b , { -1, -1, -1, -1 } , { 0, 0, 0, 0 } >
>
>where the operands of the comparison are same except for the sign (it's
>vector_types_compatible_elements_p that says that). But GIMPLE doesn't
>like the difference in the sign.
>
>As I read the discussion in the PR, the way forward here is to perform
>signed -> unsigned conversions. So that is what I do in the following.
>
>Bootstrapped/regtested on x86_64-linux and ppc64le-linux, ok for trunk?
Looks good from a middle-end perspective.
Richard.
>2016-01-13 Marek Polacek <pola...@redhat.com>
>
> PR c/68062
> * c-typeck.c (build_binary_op) [EQ_EXPR, GE_EXPR]: Promote operand
> to unsigned, if needed.
>
> * typeck.c (cp_build_binary_op): Promote operand to unsigned, if
> needed.
>
> * c-c++-common/vector-compare-4.c: New test.
>
>diff --git gcc/c/c-typeck.c gcc/c/c-typeck.c
>index 952041b..b646451 100644
>--- gcc/c/c-typeck.c
>+++ gcc/c/c-typeck.c
>@@ -11048,6 +11048,18 @@ build_binary_op (location_t location, enum
>tree_code code,
> return error_mark_node;
> }
>
>+ /* It's not precisely specified how the usual arithmetic
>+ conversions apply to the vector types. Here, we use
>+ the unsigned type if one of the operands is signed and
>+ the other one is unsigned. */
>+ if (TYPE_UNSIGNED (type0) != TYPE_UNSIGNED (type1))
>+ {
>+ if (!TYPE_UNSIGNED (type0))
>+ op0 = build1 (VIEW_CONVERT_EXPR, type1, op0);
>+ else
>+ op1 = build1 (VIEW_CONVERT_EXPR, type0, op1);
>+ }
>+
> /* Always construct signed integer vector type. */
> intt = c_common_type_for_size (GET_MODE_BITSIZE
> (TYPE_MODE (TREE_TYPE (type0))), 0);
>@@ -11201,6 +11213,18 @@ build_binary_op (location_t location, enum
>tree_code code,
> return error_mark_node;
> }
>
>+ /* It's not precisely specified how the usual arithmetic
>+ conversions apply to the vector types. Here, we use
>+ the unsigned type if one of the operands is signed and
>+ the other one is unsigned. */
>+ if (TYPE_UNSIGNED (type0) != TYPE_UNSIGNED (type1))
>+ {
>+ if (!TYPE_UNSIGNED (type0))
>+ op0 = build1 (VIEW_CONVERT_EXPR, type1, op0);
>+ else
>+ op1 = build1 (VIEW_CONVERT_EXPR, type0, op1);
>+ }
>+
> /* Always construct signed integer vector type. */
> intt = c_common_type_for_size (GET_MODE_BITSIZE
> (TYPE_MODE (TREE_TYPE (type0))), 0);
>diff --git gcc/cp/typeck.c gcc/cp/typeck.c
>index 472b41b..ffa9ed4 100644
>--- gcc/cp/typeck.c
>+++ gcc/cp/typeck.c
>@@ -4813,6 +4813,18 @@ cp_build_binary_op (location_t location,
> return error_mark_node;
> }
>
>+ /* It's not precisely specified how the usual arithmetic
>+ conversions apply to the vector types. Here, we use
>+ the unsigned type if one of the operands is signed and
>+ the other one is unsigned. */
>+ if (TYPE_UNSIGNED (type0) != TYPE_UNSIGNED (type1))
>+ {
>+ if (!TYPE_UNSIGNED (type0))
>+ op0 = build1 (VIEW_CONVERT_EXPR, type1, op0);
>+ else
>+ op1 = build1 (VIEW_CONVERT_EXPR, type0, op1);
>+ }
>+
> /* Always construct signed integer vector type. */
> intt = c_common_type_for_size (GET_MODE_BITSIZE
> (TYPE_MODE (TREE_TYPE (type0))), 0);
>diff --git gcc/testsuite/c-c++-common/vector-compare-4.c
>gcc/testsuite/c-c++-common/vector-compare-4.c
>index e69de29..14faf04 100644
>--- gcc/testsuite/c-c++-common/vector-compare-4.c
>+++ gcc/testsuite/c-c++-common/vector-compare-4.c
>@@ -0,0 +1,41 @@
>+/* PR c/68062 */
>+/* { dg-do compile } */
>+
>+typedef signed char __attribute__ ((vector_size (4))) v4qi;
>+typedef unsigned char __attribute__ ((vector_size (4))) uv4qi;
>+typedef signed int __attribute__ ((vector_size (4 * __SIZEOF_INT__)))
>v4si;
>+typedef unsigned int __attribute__ ((vector_size (4 *
>__SIZEOF_INT__))) uv4si;
>+
>+v4qi
>+fn1 (void)
>+{
>+ v4qi a = { 1, 2, 3, 4 };
>+ uv4qi b = { 4, 3, 2, 1 };
>+ v4qi v = { 0, 0, 0, 0 };
>+
>+ v += (a == b);
>+ v += (a != b);
>+ v += (a >= b);
>+ v += (a <= b);
>+ v += (a > b);
>+ v += (a < b);
>+
>+ return v;
>+}
>+
>+v4si
>+fn2 (void)
>+{
>+ v4si a = { 1, 2, 3, 4 };
>+ uv4si b = { 4, 3, 2, 1 };
>+ v4si v = { 0, 0, 0, 0 };
>+
>+ v += (a == b);
>+ v += (a != b);
>+ v += (a >= b);
>+ v += (a <= b);
>+ v += (a > b);
>+ v += (a < b);
>+
>+ return v;
>+}
>
> Marek
