C++ 为什么在不同的上下文中调用 pow() 有时会产生不同的结果

我想知道如何解释我今天观察到的浮点差异.这是一个错误，还是触发了一些未定义的行为？以下是代码.我希望理解这种行为，这样我就可以使它在编译器、处理器和平台上保持一致.

#include <stdio.h>
#include <math.h>

#define BUG_ON_GCC
#ifdef BUG_ON_GCC
__attribute__((noinline))
#else
inline __attribute__((always_inline))
#endif
double safe_pow(
  double x,
  double y)
{
  printf("x = %llx\n", *((unsigned long long *)&x));
  printf("y = %llx\n", *((unsigned long long *)&y));
  double result = pow(x, y);
  printf("r = %llx\n", *((unsigned long long *)&result));
  return result;
}

int main() {
  printf("%ld\n", sizeof(pow(15.034465284692086, 3.466120406090667)));
  printf("%.24f\n", pow(15.034465284692086, 3.466120406090667));
  printf("%ld\n", sizeof(safe_pow(15.034465284692086, 3.466120406090667)));
  printf("%.24f\n", safe_pow(15.034465284692086, 3.466120406090667));
  return 0;
}

以及我在Linux VM中使用Clang所看到的结果:

8
12020.670425990641888347454369
8
x = 402e11a56f0d331e
y = 400bba9d55e142e0
r = 40c77a55d084d419
12020.670425990641888347454369

在同一个VM中使用GCC:

8
12020.670425990643707336857915
8
x = 402e11a56f0d331e
y = 400bba9d55e142e0
r = 40c77a55d084d419
12020.670425990641888347454369

在同一台计算机上使用MacOS上的Clang:

8
12020.670425990643707336857915
8
x = 402e11a56f0d331e
y = 400bba9d55e142e0
r = 40c77a55d084d41a
12020.670425990643707336857915