http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58493
Bug ID: 58493
Summary: loop is not correctly optimized with O3 and AVX
Product: gcc
Version: 4.8.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: vgrebinski at gmail dot com
Host: Debian 3.9.6-1 x86_64 GNU/Linux; Intel(R) Xeon(R) CPU
E5-2687W
Target: amd64/sandy bridge
The following example (simplified from a proprietary numerical quant library
issue) shows that AVX loop optimizer mis-compiles code below with both
gcc-4.8.1 and gcc-4.7.3 : the first eight weights[2*i+1] values are wrong.
Happens only when AVX instructions are available (i.e. "-O3 -mavx"). Does not
happen with "O2" or sse2.
The fact that points[] and weights[] assignments are interleaved is important
(no bug otherwise).
Obviously, example has to be run on a system that supports AVX.
///////////////////////// start
#include <iostream>
#include <vector>
using namespace std;
void omb(size_t n, vector<double>& points, vector<double>& weights)
{
points.resize(n);
weights.resize(n);
for(int i=0;i<n/2;i++)
{
points[2*i] = .7;
weights[2*i]= 5.;
points[2*i+1] = -.7;
weights[2*i+1]=weights[2*i]; // mis-comiled
}
}
int main()
{
vector<double> p,w;
omb(18, p, w);
for(size_t i=0; i!= p.size(); ++i)
cout << "i= " << i << " p= " << p[i] << " w= " << w[i] << endl;
}
/////////////////////////////// end
With gcc-4.7.3:
> g++-4.7 --version && g++-4.7 -O3 -mavx bug.cc && ./a.out
g++-4.7 (Debian 4.7.3-6) 4.7.3
Copyright (C) 2012 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
i= 0 p= 0.7 w= 5
i= 1 p= -0.7 w= 0 ## BUG, should be 5
i= 2 p= 0.7 w= 5
i= 3 p= -0.7 w= 0 ## BUG
i= 4 p= 0.7 w= 5
i= 5 p= -0.7 w= 0 ## BUG
i= 6 p= 0.7 w= 5
i= 7 p= -0.7 w= 0 ## BUG
i= 8 p= 0.7 w= 5
i= 9 p= -0.7 w= 0 ## BUG
i= 10 p= 0.7 w= 5
i= 11 p= -0.7 w= 0 ## BUG
i= 12 p= 0.7 w= 5
i= 13 p= -0.7 w= 0 ## BUG
i= 14 p= 0.7 w= 5
i= 15 p= -0.7 w= 0 ## BUG
i= 16 p= 0.7 w= 5
i= 17 p= -0.7 w= 5 ## right value
With gcc-4.8:
> g++-4.8 --version && g++-4.8 -O3 -mavx bug.cc && ./a.out
g++-4.8 (Debian 4.8.1-10) 4.8.1
Copyright (C) 2013 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
i= 0 p= 0.7 w= 5
i= 1 p= -0.7 w= 0 # BUG, as well as the next 7 odd-numbered entries
i= 2 p= 0.7 w= 5
i= 3 p= -0.7 w= 0
i= 4 p= 0.7 w= 5
i= 5 p= -0.7 w= 0
i= 6 p= 0.7 w= 5
i= 7 p= -0.7 w= 0
i= 8 p= 0.7 w= 5
i= 9 p= -0.7 w= 0
i= 10 p= 0.7 w= 5
i= 11 p= -0.7 w= 0
i= 12 p= 0.7 w= 5
i= 13 p= -0.7 w= 0
i= 14 p= 0.7 w= 5
i= 15 p= -0.7 w= 0
i= 16 p= 0.7 w= 5
i= 17 p= -0.7 w= 5
Same, but SSE2, not AVX: initialized correctly.
> g++-4.8 --version && g++-4.8 -O3 -msse2 bug.cc && ./a.out
g++-4.8 (Debian 4.8.1-10) 4.8.1
Copyright (C) 2013 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
i= 0 p= 0.7 w= 5
i= 1 p= -0.7 w= 5 # OK
i= 2 p= 0.7 w= 5
i= 3 p= -0.7 w= 5
i= 4 p= 0.7 w= 5
i= 5 p= -0.7 w= 5
i= 6 p= 0.7 w= 5
i= 7 p= -0.7 w= 5
i= 8 p= 0.7 w= 5
i= 9 p= -0.7 w= 5
i= 10 p= 0.7 w= 5
i= 11 p= -0.7 w= 5
i= 12 p= 0.7 w= 5
i= 13 p= -0.7 w= 5
i= 14 p= 0.7 w= 5
i= 15 p= -0.7 w= 5
i= 16 p= 0.7 w= 5
i= 17 p= -0.7 w= 5
Now gcc-4.6 & AVX : no bug
> g++-4.6 --version && g++-4.6 -O3 -mavx bug.cc && ./a.out
g++-4.6 (Debian 4.6.4-4) 4.6.4
Copyright (C) 2011 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
i= 0 p= 0.7 w= 5
i= 1 p= -0.7 w= 5 # OK
i= 2 p= 0.7 w= 5
i= 3 p= -0.7 w= 5
i= 4 p= 0.7 w= 5
i= 5 p= -0.7 w= 5
i= 6 p= 0.7 w= 5
i= 7 p= -0.7 w= 5
i= 8 p= 0.7 w= 5
i= 9 p= -0.7 w= 5
i= 10 p= 0.7 w= 5
i= 11 p= -0.7 w= 5
i= 12 p= 0.7 w= 5
i= 13 p= -0.7 w= 5
i= 14 p= 0.7 w= 5
i= 15 p= -0.7 w= 5
i= 16 p= 0.7 w= 5
i= 17 p= -0.7 w= 5
Regards,
Vladimir
