https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88998
Bug ID: 88998
Summary: bad codegen with mmx instructions for unordered_map
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: barry.revzin at gmail dot com
Target Milestone: ---
This program should be valid:
#include <cassert>
#include <unordered_map>
#include <x86intrin.h>
[[gnu::noinline]]
double prepare(int a, int b)
{
__m128i is = _mm_setr_epi32(a, b, 0, 0);
__m128d ds =_mm_cvtepi32_pd(is);
return ds[0] + ds[1];
}
int main(int, char**) {
double d = prepare(1, 2);
std::unordered_map<int, int> m;
m.insert({0, 0});
m.insert({1, 1});
assert(m.load_factor() <= m.max_load_factor());
return d;
}
But if I use MMX instructions, the assertion triggers:
$ g++ -std=c++11 -march=haswell -mtune=haswell -mavx -O3 foo.cxx && ./a.out
a.out: foo.cxx:19: int main(int, char**): Assertion `m.load_factor() <=
m.max_load_factor()' failed.
Aborted (core dumped)
Whereas it does not if I explicitly remove them:
$ g++ -std=c++11 -march=haswell -mtune=haswell -mavx -mno-mmx -O3 foo.cxx &&
./a.out
$
Additionally, adding a call to _mm_empty() right before the return statement in
prepare() fixes the issue.
If you look at the load_factor(), it returns 2 (and then 3 on the next insert,
etc.)
