https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120895
Bug ID: 120895
Summary: AVX data types default alignment is not correct
Product: gcc
Version: 13.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: david.partridge at perdrix dot co.uk
Target Milestone: ---
Created attachment 61767
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=61767&action=edit
test case demonstrating problem
The alignment *requirement* of __m512 is 64 bytes, and as of C++17, all
elements of a std::vector<__m512> *must* be 64 byte aligned.
The compiler is setting a required alignment of 16 bytes, not 64 bytes, but the
type requires alignment of 64 bytes. This is NOT a case where it has better
performance when aligned to 64 bytes. It is *necessary* to avoid SIGSEGV faults
in the intrinsics.
Similar comments apply to most other SIMD datatypes.
Using g++ (Ubuntu 13.1.0-8ubuntu1~22.04) 13.1.0
The elements of such vectors aren't even aligned on 32 byte boundaries!!!
System is x86_64
See attached file for test case
When I run the test case I get:
amonra@styx:~/m512$ ./a.out
std::alignment_of_v<__m512> is: 16
alignof(__m512) is: 16
__alignof__(__m512) is: 64
Address of data%64: 0
Address1%64: 0 Address2%64: 16
a.out: m512.cpp:35: int main(int, char**): Assertion `align2 == 0' failed.
Aborted (core dumped)
amonra@styx:~/m512$
It all works fine on MSVC (VS2022)
I know it is possible to work round this by using e.g.:
#include <boost/align/aligned_allocator.hpp>
:
typedef std::vector<VectorElementType,
boost::alignment::aligned_allocator<VectorElementType, 64> > VectorType;
but that should not be necessary.
Please ensure that all SIMD data types are correctly aligned
Thanks
David
