[Bug target/56027] New: ldmxcsr permuted with asm

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56027



             Bug #: 56027

           Summary: ldmxcsr permuted with asm

    Classification: Unclassified

           Product: gcc

           Version: 4.8.0

            Status: UNCONFIRMED

          Severity: normal

          Priority: P3

         Component: target

        AssignedTo: unassig...@gcc.gnu.org

        ReportedBy: gli...@gcc.gnu.org

            Target: x86_64-linux-gnu





Hello,



up to now, I had the impression that the use of a function like "opaque" below

was enough to prevent bad optimizations. However, in the following program,

compiled with g++ -O1 -frounding-math, the assertion fails. sqrtsd and *mxcsr

seem to get permuted (possibly in the loop2_* RTL passes). Defining macro V1 or

V2 seems to fix it, but it is slower and I don't know if it is completely safe.

Marking the asm volatile also works (and should have less impact on speed), but

again I don't know how safe that is.



#include <xmmintrin.h>

#include <assert.h>

#include <fenv.h>

inline double opaque(double x){

#ifndef V1

  asm("":"+xm"(x));

#else

  asm("":"+m"(x));

#endif

  return x;

}

inline double f(double x){

  double ret = __builtin_sqrt(opaque(x));

  return opaque(ret);

}

int main(){

  double d=2;

  double l,h;

  for(int i=0;i<100;++i){

#ifndef V2

  _MM_SET_ROUNDING_MODE(_MM_ROUND_DOWN);

#else

  fesetround(FE_DOWNWARD);

#endif

  l=f(d);

#ifndef V2

  _MM_SET_ROUNDING_MODE(_MM_ROUND_UP);

#else

  fesetround(FE_UPWARD);

#endif

  h=f(d);

  }

  assert(l!=h);

}