c, c++: Fix up excess precision handling of scalar_to_vector conversion [PR107358]
Checks
Commit Message
Hi!
As mentioned earlier in the C++ excess precision support mail, the following
testcase is broken with excess precision both in C and C++ (though just in C++
it was triggered in real-world code).
scalar_to_vector is called in both FEs after the excess precision promotions
(or stripping of EXCESS_PRECISION_EXPR), so we can then get invalid
diagnostics that say float vector + float involves truncation (on ia32
from long double to float).
The following patch fixes that by calling scalar_to_vector on the operands
before the excess precision promotions, let scalar_to_vector just do the
diagnostics (it does e.g. fold_for_warn so it will fold
EXCESS_PRECISION_EXPR around REAL_CST to constants etc.) but will then
do the actual conversions using the excess precision promoted operands
(so say if we have vector double + (float + float) we don't actually do
vector double + (float) ((long double) float + (long double) float)
but
vector double + (double) ((long double) float + (long double) float)
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2022-10-24 Jakub Jelinek <jakub@redhat.com>
PR c++/107358
c/
* c-typeck.cc (build_binary_op): Pass operands before excess precision
promotions to scalar_to_vector call.
cp/
* typeck.cc (cp_build_binary_op): Pass operands before excess precision
promotions to scalar_to_vector call.
testsuite/
* c-c++-common/pr107358.c: New test.
* g++.dg/cpp1y/pr68180.C: Remove -fexcess-precision=fast from
dg-options.
Jakub
Comments
On 10/24/22 01:19, Jakub Jelinek via Gcc-patches wrote:
> Hi!
>
> As mentioned earlier in the C++ excess precision support mail, the following
> testcase is broken with excess precision both in C and C++ (though just in C++
> it was triggered in real-world code).
> scalar_to_vector is called in both FEs after the excess precision promotions
> (or stripping of EXCESS_PRECISION_EXPR), so we can then get invalid
> diagnostics that say float vector + float involves truncation (on ia32
> from long double to float).
>
> The following patch fixes that by calling scalar_to_vector on the operands
> before the excess precision promotions, let scalar_to_vector just do the
> diagnostics (it does e.g. fold_for_warn so it will fold
> EXCESS_PRECISION_EXPR around REAL_CST to constants etc.) but will then
> do the actual conversions using the excess precision promoted operands
> (so say if we have vector double + (float + float) we don't actually do
> vector double + (float) ((long double) float + (long double) float)
> but
> vector double + (double) ((long double) float + (long double) float)
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2022-10-24 Jakub Jelinek <jakub@redhat.com>
>
> PR c++/107358
> c/
> * c-typeck.cc (build_binary_op): Pass operands before excess precision
> promotions to scalar_to_vector call.
> cp/
> * typeck.cc (cp_build_binary_op): Pass operands before excess precision
> promotions to scalar_to_vector call.
> testsuite/
> * c-c++-common/pr107358.c: New test.
> * g++.dg/cpp1y/pr68180.C: Remove -fexcess-precision=fast from
> dg-options.
OK
jeff
On 10/24/22 03:19, Jakub Jelinek wrote:
> Hi!
>
> As mentioned earlier in the C++ excess precision support mail, the following
> testcase is broken with excess precision both in C and C++ (though just in C++
> it was triggered in real-world code).
> scalar_to_vector is called in both FEs after the excess precision promotions
> (or stripping of EXCESS_PRECISION_EXPR), so we can then get invalid
> diagnostics that say float vector + float involves truncation (on ia32
> from long double to float).
>
> The following patch fixes that by calling scalar_to_vector on the operands
> before the excess precision promotions, let scalar_to_vector just do the
> diagnostics (it does e.g. fold_for_warn so it will fold
> EXCESS_PRECISION_EXPR around REAL_CST to constants etc.) but will then
> do the actual conversions using the excess precision promoted operands
> (so say if we have vector double + (float + float) we don't actually do
> vector double + (float) ((long double) float + (long double) float)
> but
> vector double + (double) ((long double) float + (long double) float)
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
OK on Wednesday if Joseph doesn't object.
> 2022-10-24 Jakub Jelinek <jakub@redhat.com>
>
> PR c++/107358
> c/
> * c-typeck.cc (build_binary_op): Pass operands before excess precision
> promotions to scalar_to_vector call.
> cp/
> * typeck.cc (cp_build_binary_op): Pass operands before excess precision
> promotions to scalar_to_vector call.
> testsuite/
> * c-c++-common/pr107358.c: New test.
> * g++.dg/cpp1y/pr68180.C: Remove -fexcess-precision=fast from
> dg-options.
>
> --- gcc/c/c-typeck.cc.jj 2022-10-14 09:35:56.199990261 +0200
> +++ gcc/c/c-typeck.cc 2022-10-22 17:54:24.378839301 +0200
> @@ -11995,8 +11995,8 @@ build_binary_op (location_t location, en
> if ((gnu_vector_type_p (type0) && code1 != VECTOR_TYPE)
> || (gnu_vector_type_p (type1) && code0 != VECTOR_TYPE))
> {
> - enum stv_conv convert_flag = scalar_to_vector (location, code, op0, op1,
> - true);
> + enum stv_conv convert_flag = scalar_to_vector (location, code, orig_op0,
> + orig_op1, true);
>
> switch (convert_flag)
> {
> --- gcc/cp/typeck.cc.jj 2022-10-20 13:54:22.535670240 +0200
> +++ gcc/cp/typeck.cc 2022-10-22 17:56:58.589715301 +0200
> @@ -5191,6 +5191,8 @@ cp_build_binary_op (const op_location_t
>
> orig_type0 = type0 = TREE_TYPE (op0);
> orig_type1 = type1 = TREE_TYPE (op1);
> + tree non_ep_op0 = op0;
> + tree non_ep_op1 = op1;
>
> /* The expression codes of the data types of the arguments tell us
> whether the arguments are integers, floating, pointers, etc. */
> @@ -5303,8 +5305,9 @@ cp_build_binary_op (const op_location_t
> if ((gnu_vector_type_p (type0) && code1 != VECTOR_TYPE)
> || (gnu_vector_type_p (type1) && code0 != VECTOR_TYPE))
> {
> - enum stv_conv convert_flag = scalar_to_vector (location, code, op0, op1,
> - complain & tf_error);
> + enum stv_conv convert_flag
> + = scalar_to_vector (location, code, non_ep_op0, non_ep_op1,
> + complain & tf_error);
>
> switch (convert_flag)
> {
> --- gcc/testsuite/c-c++-common/pr107358.c.jj 2022-10-22 18:46:59.390375310 +0200
> +++ gcc/testsuite/c-c++-common/pr107358.c 2022-10-22 18:01:52.973660719 +0200
> @@ -0,0 +1,30 @@
> +/* PR c++/107358 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fexcess-precision=standard" } */
> +
> +typedef float __attribute__((vector_size (4 * sizeof (float)))) A;
> +typedef double __attribute__((vector_size (2 * sizeof (double)))) B;
> +
> +void
> +foo (A *x)
> +{
> + *x = *x - 124.225514990f;
> +}
> +
> +void
> +bar (A *x, float y)
> +{
> + *x = *x - y;
> +}
> +
> +void
> +baz (B *x)
> +{
> + *x = *x + 124.225514990f;
> +}
> +
> +void
> +qux (B *x, double y)
> +{
> + *x = *x + y;
> +}
> --- gcc/testsuite/g++.dg/cpp1y/pr68180.C.jj 2022-10-14 09:28:28.339159477 +0200
> +++ gcc/testsuite/g++.dg/cpp1y/pr68180.C 2022-10-22 17:59:07.012946513 +0200
> @@ -1,6 +1,6 @@
> // PR c++/68180
> // { dg-do compile { target c++14 } }
> -// { dg-additional-options "-Wno-psabi -fexcess-precision=fast" }
> +// { dg-additional-options "-Wno-psabi" }
>
> typedef float __attribute__( ( vector_size( 16 ) ) ) float32x4_t;
> constexpr float32x4_t fill(float x) {
>
> Jakub
>
@@ -11995,8 +11995,8 @@ build_binary_op (location_t location, en
if ((gnu_vector_type_p (type0) && code1 != VECTOR_TYPE)
|| (gnu_vector_type_p (type1) && code0 != VECTOR_TYPE))
{
- enum stv_conv convert_flag = scalar_to_vector (location, code, op0, op1,
- true);
+ enum stv_conv convert_flag = scalar_to_vector (location, code, orig_op0,
+ orig_op1, true);
switch (convert_flag)
{
@@ -5191,6 +5191,8 @@ cp_build_binary_op (const op_location_t
orig_type0 = type0 = TREE_TYPE (op0);
orig_type1 = type1 = TREE_TYPE (op1);
+ tree non_ep_op0 = op0;
+ tree non_ep_op1 = op1;
/* The expression codes of the data types of the arguments tell us
whether the arguments are integers, floating, pointers, etc. */
@@ -5303,8 +5305,9 @@ cp_build_binary_op (const op_location_t
if ((gnu_vector_type_p (type0) && code1 != VECTOR_TYPE)
|| (gnu_vector_type_p (type1) && code0 != VECTOR_TYPE))
{
- enum stv_conv convert_flag = scalar_to_vector (location, code, op0, op1,
- complain & tf_error);
+ enum stv_conv convert_flag
+ = scalar_to_vector (location, code, non_ep_op0, non_ep_op1,
+ complain & tf_error);
switch (convert_flag)
{
@@ -0,0 +1,30 @@
+/* PR c++/107358 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -fexcess-precision=standard" } */
+
+typedef float __attribute__((vector_size (4 * sizeof (float)))) A;
+typedef double __attribute__((vector_size (2 * sizeof (double)))) B;
+
+void
+foo (A *x)
+{
+ *x = *x - 124.225514990f;
+}
+
+void
+bar (A *x, float y)
+{
+ *x = *x - y;
+}
+
+void
+baz (B *x)
+{
+ *x = *x + 124.225514990f;
+}
+
+void
+qux (B *x, double y)
+{
+ *x = *x + y;
+}
@@ -1,6 +1,6 @@
// PR c++/68180
// { dg-do compile { target c++14 } }
-// { dg-additional-options "-Wno-psabi -fexcess-precision=fast" }
+// { dg-additional-options "-Wno-psabi" }
typedef float __attribute__( ( vector_size( 16 ) ) ) float32x4_t;
constexpr float32x4_t fill(float x) {