[dmd-internals] [D-Programming-Language/dmd] 48ed15: Remove import of std.random:rand from test22

[Walter Bright]

Don, could you add this to the bug report, please?

On 8/23/2011 5:46 AM, Don Clugston wrote:
>
>>> The backend does spill the x87 stack when it overflows. I suspect the
>>> problem may lie in the scheduler. This is easily checked by disabling the
>>> scheduler and trying it.
>> You're right, it's the scheduler. It loads 8 values on the x87 stack,
>> then does fdecstp.
>> Then, the 9th load lands on top of an existing value, causing a stack
>> overflow exception. The loaded value gets turned into a NaN.
>> I'll try to track it down further.
> OK, here's the problem. The scheduler keeps track of the number of
> used x87 registers via the "fpustackused" variable.
> Each instruction has a "fpuadjust" value which says how many x87
> registers it uses.
> But, the problem is CALL instructions. They have fpuadjust = 0.
> But, a function returning float returns it in ST(0) --- so fpuadjust
> should be +1.
> And if it's a complex result, fpuadjust should be +2.
> I think it should be possible to put an assert in cgsched.c, line 2491:
>
>              if (tbl[j])
>              {
>                  fpu += tbl[j]->fpuadjust;
> + assert(fpu>= 0);
>                  if (fpu>= 8)                   // if FPU stack overflow
>
> but this would fail at the moment, because the total goes negative
> after a call followed by an FSTP.
> So it needs to distinguish between each type of call, depending on how
> it affects the FPU stack.
> I think the bug lies in funccall() in cod1.c; it should be setting
> something in 'code' to record how many x87 registers are returned.
>
> Then, there'll be another minor problem: the number of FPU values CAN
> reach 8. This is because of the code in push87() in cg87.c, which does
> a
> fdecstp; fstp; when the stack is full. So the condition above will
> need to change to:
>                  if (fpu>  8)                   // if FPU stack overflow
>
> This is too invasive a change for me to make and test.
>