diff options
| -rw-r--r-- | Eigen/src/Core/util/Memory.h | 5 |
1 files changed, 3 insertions, 2 deletions
diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index b26afda64..3d0994415 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -751,9 +751,10 @@ public: # if defined(__PIC__) && defined(__i386__) // Case for x86 with PIC # define EIGEN_CPUID(abcd,func,id) \ - __asm__ __volatile__ ("xchgl %%ebx, %%esi;cpuid; xchgl %%ebx,%%esi": "=a" (abcd[0]), "=S" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id)); + __asm__ __volatile__ ("xchgl %%ebx, %k1;cpuid; xchgl %%ebx,%k1": "=a" (abcd[0]), "=&r" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id)); # elif defined(__PIC__) && defined(__x86_64__) - // Case for x64 with PIC + // Case for x64 with PIC. In theory this is only a problem with recent gcc and with medium or large code model, not with the default small code model. + // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway. # define EIGEN_CPUID(abcd,func,id) \ __asm__ __volatile__ ("xchg{q}\t{%%}rbx, %q1; cpuid; xchg{q}\t{%%}rbx, %q1": "=a" (abcd[0]), "=&r" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id)); # else |