People have posted the asm now:

https://mail.python.org/pipermail/python-list/2015-July/693509.html


The fmull result appears to be the same. The good gcc versions set
the control word to "truncate" and then use fistpl (single rounding
with the correct mode):

  fmull  -0x10(%ebp)
  fnstcw -0x1a(%ebp)
  movzwl -0x1a(%ebp),%eax
  mov    $0xc,%ah
  mov    %ax,-0x1c(%ebp)
  fldcw  -0x1c(%ebp)
  fistpl -0x20(%ebp)


The bad versions dump the result of fmull to memory, using
"round-to-nearest" and *then* attempt to truncate with
cvttsd2si, which is too late.
 
  fmull  0x18(%esp)
​  fstpl  0x8(%esp)
  cvttsd2si 0x8(%esp),%eax


I suspect that there are passages in various standards that allow
both behaviors. ;)