http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54498
Andrew Pinski <pinskia at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Status|UNCONFIRMED |NEW Keywords| |wrong-code Last reconfirmed| |2012-09-05 Component|c++ |tree-optimization Ever Confirmed|0 |1 Summary|incorrect code generation |[4.7/4.8 Regression] |from g++ -O on x86_64 |incorrect code generation | |from g++ -O Target Milestone|--- |4.7.2 --- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> 2012-09-05 22:21:58 UTC --- Confirmed, here is the source unincluded: #include <complex> #include <cstdio> using namespace std; class bar_src { public: bar_src() : next(0) {} virtual ~bar_src() { delete next; } bar_src *next; }; class foo_src : public bar_src { public: foo_src(double f, double fwidth, double s = 5.0); virtual ~foo_src() {} private: double freq, width, peak_time, cutoff; }; foo_src::foo_src(double f, double fwidth, double s) { freq = f; width = 1/fwidth; cutoff = s*width; peak_time = cutoff; } complex<double> do_ft2(int i) __attribute__ ((noinline)); complex<double> do_ft2(int i) { return i == 0 ? complex<double>(-491.697,887.05) : complex<double>(-491.692,887.026); } void foo(void) { complex<double> prev_ft = 0.0, ft = 0.0; for (int i=0; i < 2; i++) { prev_ft = ft; { foo_src src(1.0, 1.0 / 20); ft = do_ft2(i); printf("ft (it = %d) = %g%+gi\n", i, real(ft), imag(ft)); } if (i > 0) printf("abs(ft - prev_ft) = %g\n", abs(ft - prev_ft)); } } int main(void) { foo(); return 0; } --- CUT --- The problem comes from FRE, I think due to how it handles look though copies. It happens on more than just x86_64.