[Bug rtl-optimization/69941] bogus zero_extend transformation in postreload on aarch64

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69941

--- Comment #3 from ktkachov at gcc dot gnu.org ---
Sorry, here's the standalone testcase using __builtin_abort:


int a = 0;
int b = 0;
int c = 0;
int e = 0;
int f = 0;
int *g = &e;

int fn1() { return b ? a : b; }

int main() {
  int h = fn1() <= 0x8000000000000000ULL; // h = 1;

  int k = f; // k = 0;

  long i = h ? k : k / h; // i = 0;

  long l = (unsigned short)(i - 0x1800); // l = 0xe800

  i = l ? l : c; // i = 0xe800;

  *g = i; // *g = 0xe800; e = 0xe800;

  unsigned char result = e >> 9; // result = 0x74;

  if ((int)result != 0x74)
    __builtin_abort ();
  return 0;
}

This aborts on aarch64 at -O1, and doesn't abort at -O0.
As mentioned before, this bug goes latent with r232540, so you need a revision
before that