[llvm-bugs] [Bug 47674] New: Clang discards attributes aligned and may_alias for typedefs passed as template arguments
via llvm-bugs
llvm-bugs at lists.llvm.org
Mon Sep 28 17:37:06 PDT 2020
https://bugs.llvm.org/show_bug.cgi?id=47674
Bug ID: 47674
Summary: Clang discards attributes aligned and may_alias for
typedefs passed as template arguments
Product: clang
Version: 10.0
Hardware: All
OS: Linux
Status: NEW
Severity: normal
Priority: P
Component: C++
Assignee: unassignedclangbugs at nondot.org
Reporter: mte.zych at gmail.com
CC: blitzrakete at gmail.com, dgregor at apple.com,
erik.pilkington at gmail.com, llvm-bugs at lists.llvm.org,
richard-llvm at metafoo.co.uk
Hello!
Clang discards aligned attribute, applied on a typdef,
when it's passed as a template argument.
Compiler Expolorer:
Clang -> https://godbolt.org/z/x375GT
C++ Source Code:
#include <iostream>
typedef float vec __attribute__((vector_size(8)));
typedef float fp __attribute__((aligned(16)));
template <typename t> struct identity { typedef t type; };
int main ()
{
std::cout << sizeof(typename identity<vec>::type) << std::endl;
std::cout << sizeof(vec ) << std::endl;
std::cout << alignof(typename identity<fp>::type) << std::endl;
std::cout << alignof(fp ) << std::endl;
}
Program Output:
8
8
4
16
The above program shows that alignment of the fp typedef changes,
after it's been passed through the identity meta-function - it's 4-bytes
instead of expected 16-bytes.
What's interesting is not all type attributes are discarded - the vector_size
attribute is preserved after being passed through the identity meta-function.
This behavior is required, since removal of the vector_size attribute
would be a semantic change of the vec type, affecting even its size,
because the vec type would represent a single float,
instead of a vector of 2 floats.
The same could be said about the aligned attribute - discarding it
is also a semantic change, since alignment is a fundamental property of a type,
affecting among others code generation,
that is, two types are not equivalent if they have different alignment.
This is the reason why I argue that, passing a typedef as a template argument
should preserve its aligned attribute, instead of discarding it.
Moreover, the Intel C++ compiler implements this behavior correctly.
Compiler Expolorer:
ICC -> https://godbolt.org/z/9vr9se
Program Output:
8
8
16
16
The issue described above doesn't apply only to the aligned type attribute,
but also to the may_alias type attribute.
Compiler Expolorer:
Clang -> https://godbolt.org/z/zYEz78
C++ Source Code:
#include <iostream>
#include <limits>
typedef float fp __attribute__((may_alias));
template <typename T> struct identity { typedef T type; };
static_assert( sizeof(float) == sizeof(int) , "");
static_assert(alignof(float) == alignof(int) , "");
static_assert(std::numeric_limits<float>::is_iec559, "");
bool can_alias_float ()
{
const auto fn = [] (float *f, int *i) -> int
{
*i = 0x1;
*f = 2.0f; // In ieee754 bin repr of 2.0f is 0x40000000.
return *i;
};
int val; // Casting int* to float* is UB!
val = fn(reinterpret_cast<float*>(&val), &val);
return val == 0x40000000;
}
bool can_alias_fp ()
{
const auto fn = [] (fp *f, int *i) -> int
{
*i = 0x1;
*f = 2.0f; // In ieee754 bin repr of 2.0f is 0x40000000.
return *i;
};
int val; // Casting int* to fp* is OK, due to attribute may_alias.
val = fn(reinterpret_cast<fp*>(&val), &val);
return val == 0x40000000;
}
bool can_alias_identity_type_fp ()
{
const auto fn = [] (typename identity<fp>::type *f, int *i) -> int
{
*i = 0x1;
*f = 2.0f; // In ieee754 bin repr of 2.0f is 0x40000000.
return *i;
}; // Casting int* to fp* should be OK,
int val; // but the attribute may_alias is discarded, causing UB!
val = fn(reinterpret_cast<typename identity<fp>::type*>(&val), &val);
return val == 0x40000000;
}
int main ()
{
std::cout << "fp " << can_alias_fp () <<
'\n';
std::cout << "identity<fp>::type " << can_alias_identity_type_fp() <<
'\n';
std::cout << "float " << can_alias_float () <<
'\n';
}
Again, discarding attribute may_alias is a semantic change,
because aliasing rules can affect code generation.
Why this issue is important?
Well, because it prevents generic programming, via C++ templates,
using x86 SIMD types.
If we would look at definitions of x86 SIMD types, we will notice that they are
essentially typedefs with attributes vector_size and may_alias applied on them:
- immintrin.h
typedef float __m256 __attribute__((__vector_size__(32),
__may_alias__));
- emmintrin.h
typedef long long __m128i __attribute__((__vector_size__(16),
__may_alias__));
typedef double __m128d __attribute__((__vector_size__(16),
__may_alias__));
- xmmintrin.h
typedef float __m128 __attribute__((__vector_size__(16),
__may_alias__));
Note that, the may_alias attributes is required and cannot be removed:
- /usr/lib/gcc/x86_64-linux-gnu/10/include/immintrin.h
/* The Intel API is flexible enough that we must allow aliasing with other
vector types, and their scalar components. */
What's the root cause of this problem?
Well, the problem is a C++ typedef is just an alias (a new name) for the old
type,
that is, it does *not* introduce a new type.
Implementing support for attributes vector_size, aligned and may_alias
in C++ typedefs requires an opaque/strong typedef,
introducing a brand new type and storing information about applied attributes.
typedef float fp __attribute__((aligned(16)));
Think about it - a typedef introducing the fp type has to create a new type,
because even though both fp and float types represent
floating point numbers identically, the fp type is *not* the float type,
because these types have different alignment requirements.
Note that, the Intel C++ Compiler does not introduce new types for typedefs,
which have attributes aligned or may_alias applied on them.
Compiler Explorer:
ICC -> https://godbolt.org/z/MjdMqx
C++ Source Code:
#include <iostream>
typedef int vectorized_int __attribute__((vector_size(8)));
typedef int aligned_int __attribute__((aligned(16)));
typedef int aliasing_int __attribute__((may_alias));
int main ()
{
std::cout << typeid( int).name() << std::endl;
std::cout << typeid(vectorized_int).name() << std::endl;
std::cout << typeid( aligned_int).name() << std::endl;
std::cout << typeid( aliasing_int).name() << std::endl;
}
Program Output:
i
Dv2_i
i
i
However, this behavior leads to a contradiction, in which
there can exists a single type, which has 2 different alignment requirements.
Compiler Explorer:
ICC -> https://godbolt.org/z/4o9o3M
C++ Source Code:
template <class, class> struct is_same { static const auto value =
false; };
template <class T> struct is_same<T, T> { static const auto value =
true; };
typedef float fp __attribute__((aligned(16)));
template <typename first_type, typename second_type>
struct check_same
{
static_assert(is_same<first_type, second_type>::value , "");
static_assert( sizeof(first_type) == sizeof(second_type), "");
static_assert(alignof(first_type) == alignof(second_type), "");
};
int main ()
{
check_same<int, signed int> { };
check_same< fp, float> { };
}
Compilation Log:
error: static assertion failed
static_assert(alignof(first_type) == alignof(second_type), "");
^
detected during instantiation of class
"check_same<first_type, second_type> [with first_type=fp={float},
second_type=float]"
To avoid these kind of issues,
Clang could replicate the behavior of the vector_size attribute, that is,
introduce a brand new type and store information about applied attributes.
Thank you, Mateusz Zych
PS. I want to thank Ivo Kabadshow from JSC for helping me with
preparing these code samples!
--
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20200929/708aeec0/attachment.html>
More information about the llvm-bugs
mailing list