https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70970
Bug ID: 70970
Summary: Misaligned SSE with auto-vectorization
Product: gcc
Version: 5.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rcc.dark at gmail dot com
Target Milestone: ---
Created attachment 38424
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=38424&action=edit
the code
Sorry if this has been reported before. I have tested it under:
gcc version 5.3.1 20151207 (Red Hat 5.3.1-2) (GCC)
gcc version 5.2.1 20151028 (Debian 5.2.1-23)
gcc version 6.1.0 (GCC) <----- Windows
gcc version 5.2.0 (GCC) <----- Windows, MinGW-W64
The following code crashes with -std=c++14 -O3:
#include <cstdint>
#include <malloc.h>
template<typename RI>
__attribute__ ((noinline))
void symmetric_difference(RI ai, RI af, RI bi)
{
while (ai != af) {
*ai++ ^= *bi++;
}
}
int main( )
{
auto p1 = reinterpret_cast<char*>(memalign(4096, 32));
auto p2 = reinterpret_cast<char*>(memalign(4096, 32));
// _aligned_malloc under Windows
auto ai = reinterpret_cast<std::uint64_t*>(p1 + 1);
auto bi = reinterpret_cast<std::uint64_t*>(p2 + 1);
symmetric_difference(ai, ai + 64, bi);
}
It stops crashing with -O2 or if I remove the + 1 to the pointers; GDB tells me
that the problem lies within:
vmovdqa YMMWORD PTR [rbx+rcx*1],ymm0
The register rbx is not aligned and rcx = 0. It seems that it is reading with
vmovdqu but storing with vmovdqa.
