fold-const: Ignore padding bits in native_interpret_expr REAL_CST reverse verification [PR108934]
Commit Message
Hi!
In the following testcase we try to std::bit_cast a (pair of) integral
value(s) which has some non-zero bits in the place of x86 long double
(for 64-bit 16 byte type with 10 bytes actually loaded/stored by hw,
for 32-bit 12 byte) and starting with my PR104522 change we reject that
as native_interpret_expr fails on it. The PR104522 change extends what
has been done before for MODE_COMPOSITE_P (but those don't have any padding
bits) to all floating point types, because e.g. the exact x86 long double
has various bit combinations we don't support, like
pseudo-(denormals,infinities,NaNs) or unnormals. The HW handles some of
those as exceptional cases and others similarly to the non-pseudo ones.
But for the padding bits it actually doesn't load/store those bits at all,
it loads/stores 10 bytes. So, I think we should exempt the padding bits
from the reverse comparison (the native_encode_expr bits for the padding
will be all zeros), which the following patch does. For bit_cast it is
similar to e.g. ignoring padding bits if the destination is a structure
which has padding bits in there.
The change changed auto-init-4.c to how it has been behaving before the
PR105259 change, where some more VCEs can be now done.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2023-03-02 Jakub Jelinek <jakub@redhat.com>
PR c++/108934
* fold-const.cc (native_interpret_expr) <case REAL_CST>: Before memcmp
comparison copy the bytes from ptr to a temporary buffer and clearing
padding bits in there.
* gcc.target/i386/auto-init-4.c: Revert PR105259 change.
* g++.target/i386/pr108934.C: New test.
Jakub
Comments
On Thu, 2 Mar 2023, Jakub Jelinek wrote:
> Hi!
>
> In the following testcase we try to std::bit_cast a (pair of) integral
> value(s) which has some non-zero bits in the place of x86 long double
> (for 64-bit 16 byte type with 10 bytes actually loaded/stored by hw,
> for 32-bit 12 byte) and starting with my PR104522 change we reject that
> as native_interpret_expr fails on it. The PR104522 change extends what
> has been done before for MODE_COMPOSITE_P (but those don't have any padding
> bits) to all floating point types, because e.g. the exact x86 long double
> has various bit combinations we don't support, like
> pseudo-(denormals,infinities,NaNs) or unnormals. The HW handles some of
> those as exceptional cases and others similarly to the non-pseudo ones.
> But for the padding bits it actually doesn't load/store those bits at all,
> it loads/stores 10 bytes. So, I think we should exempt the padding bits
> from the reverse comparison (the native_encode_expr bits for the padding
> will be all zeros), which the following patch does. For bit_cast it is
> similar to e.g. ignoring padding bits if the destination is a structure
> which has padding bits in there.
>
> The change changed auto-init-4.c to how it has been behaving before the
> PR105259 change, where some more VCEs can be now done.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
OK.
Thanks,
Richard.
> 2023-03-02 Jakub Jelinek <jakub@redhat.com>
>
> PR c++/108934
> * fold-const.cc (native_interpret_expr) <case REAL_CST>: Before memcmp
> comparison copy the bytes from ptr to a temporary buffer and clearing
> padding bits in there.
>
> * gcc.target/i386/auto-init-4.c: Revert PR105259 change.
> * g++.target/i386/pr108934.C: New test.
>
> --- gcc/fold-const.cc.jj 2023-01-04 10:52:43.124897826 +0100
> +++ gcc/fold-const.cc 2023-03-01 16:49:14.531490482 +0100
> @@ -8873,11 +8873,13 @@ native_interpret_expr (tree type, const
> valid values that GCC can't really represent accurately.
> See PR95450. Even for other modes, e.g. x86 XFmode can have some
> bit combinationations which GCC doesn't preserve. */
> - unsigned char buf[24];
> + unsigned char buf[24 * 2];
> scalar_float_mode mode = SCALAR_FLOAT_TYPE_MODE (type);
> int total_bytes = GET_MODE_SIZE (mode);
> + memcpy (buf + 24, ptr, total_bytes);
> + clear_type_padding_in_mask (type, buf + 24);
> if (native_encode_expr (ret, buf, total_bytes, 0) != total_bytes
> - || memcmp (ptr, buf, total_bytes) != 0)
> + || memcmp (buf + 24, buf, total_bytes) != 0)
> return NULL_TREE;
> return ret;
> }
> --- gcc/testsuite/gcc.target/i386/auto-init-4.c.jj 2022-04-13 15:42:39.105365390 +0200
> +++ gcc/testsuite/gcc.target/i386/auto-init-4.c 2023-03-02 08:56:53.788029181 +0100
> @@ -15,6 +15,5 @@ long double foo()
> }
>
>
> -/* The long double init isn't expanded optimally, see PR105259. For ia32
> - it uses zero-initialization. */
> -/* { dg-final { scan-assembler-times "long\t-16843010" 3 } } */
> +/* { dg-final { scan-assembler-times "long\t-16843010" 5 { target { ! ia32 } } } } */
> +/* { dg-final { scan-assembler-times "long\t-16843010" 3 { target { ia32 } } } } */
> --- gcc/testsuite/g++.target/i386/pr108934.C.jj 2023-03-01 17:04:19.931299866 +0100
> +++ gcc/testsuite/g++.target/i386/pr108934.C 2023-03-01 17:03:27.567062785 +0100
> @@ -0,0 +1,28 @@
> +// PR c++/108934
> +// { dg-do compile { target c++11 } }
> +
> +struct S { unsigned long long a[2]; };
> +struct T { unsigned long long b[6]; };
> +struct U { unsigned long long c[2]; long double d; unsigned long long e[2]; };
> +
> +#if __SIZEOF_LONG_DOUBLE__ == 16 && __LDBL_MANT_DIG__ == 64 && __SIZEOF_LONG_LONG__ == 8
> +constexpr long double
> +foo (S x)
> +{
> + return __builtin_bit_cast (long double, x);
> +}
> +
> +constexpr S a = { 0ULL, 0xffffffffffff0000ULL };
> +constexpr long double b = foo (a);
> +static_assert (b == 0.0L, "");
> +
> +constexpr U
> +bar (T x)
> +{
> + return __builtin_bit_cast (U, x);
> +}
> +
> +constexpr T c = { 0ULL, 0ULL, 0ULL, 0xffffffffffff0000ULL, 0ULL, 0ULL };
> +constexpr U d = bar (c);
> +static_assert (d.d == 0.0L, "");
> +#endif
>
> Jakub
>
>
@@ -8873,11 +8873,13 @@ native_interpret_expr (tree type, const
valid values that GCC can't really represent accurately.
See PR95450. Even for other modes, e.g. x86 XFmode can have some
bit combinationations which GCC doesn't preserve. */
- unsigned char buf[24];
+ unsigned char buf[24 * 2];
scalar_float_mode mode = SCALAR_FLOAT_TYPE_MODE (type);
int total_bytes = GET_MODE_SIZE (mode);
+ memcpy (buf + 24, ptr, total_bytes);
+ clear_type_padding_in_mask (type, buf + 24);
if (native_encode_expr (ret, buf, total_bytes, 0) != total_bytes
- || memcmp (ptr, buf, total_bytes) != 0)
+ || memcmp (buf + 24, buf, total_bytes) != 0)
return NULL_TREE;
return ret;
}
@@ -15,6 +15,5 @@ long double foo()
}
-/* The long double init isn't expanded optimally, see PR105259. For ia32
- it uses zero-initialization. */
-/* { dg-final { scan-assembler-times "long\t-16843010" 3 } } */
+/* { dg-final { scan-assembler-times "long\t-16843010" 5 { target { ! ia32 } } } } */
+/* { dg-final { scan-assembler-times "long\t-16843010" 3 { target { ia32 } } } } */
@@ -0,0 +1,28 @@
+// PR c++/108934
+// { dg-do compile { target c++11 } }
+
+struct S { unsigned long long a[2]; };
+struct T { unsigned long long b[6]; };
+struct U { unsigned long long c[2]; long double d; unsigned long long e[2]; };
+
+#if __SIZEOF_LONG_DOUBLE__ == 16 && __LDBL_MANT_DIG__ == 64 && __SIZEOF_LONG_LONG__ == 8
+constexpr long double
+foo (S x)
+{
+ return __builtin_bit_cast (long double, x);
+}
+
+constexpr S a = { 0ULL, 0xffffffffffff0000ULL };
+constexpr long double b = foo (a);
+static_assert (b == 0.0L, "");
+
+constexpr U
+bar (T x)
+{
+ return __builtin_bit_cast (U, x);
+}
+
+constexpr T c = { 0ULL, 0ULL, 0ULL, 0xffffffffffff0000ULL, 0ULL, 0ULL };
+constexpr U d = bar (c);
+static_assert (d.d == 0.0L, "");
+#endif