crash while executing "ldmxcsr" instruction in bochscpu backend

[Micky-Thongam]

crash while executing "ldmxcsr" instruction in bochscpu backend

[0vercl0k]

Okay, we talked about the above and the suggestion didn't work. I did some digging today and I think I understand why; basically there might be a bug in bochscpu/mod.rs:

    pub unsafe fn set_state(&self, s: &State) {
        // ...
        self.set_mxcsr(s.mxcsr);
        self.set_mxcsr_mask(s.mxcsr);
    }

So editing the regs.json['mxcsr_mask'] would be a no op. What it also means is that your initial mxcsr_mask is 0x1fa1 which means both bits 1<<13 / 1<<15 are set in the new value you are loading but aren't in your mask so you get a #GP.

I'll talk to yrp604 about the above to verify that it is a bug, but to unblock you try to update regs.json['mxcsr'] to 0xBFA1 and re-launch your testcase to see if you get passed the ldmxcsr issue.

Cheers

[0vercl0k]

Okay, according to the manual, ldmxcsr #GPs when it's trying to set reserved bits. If we check Bochs' code, we see the below:

void BX_CPP_AttrRegparmN(1) BX_CPU_C::LDMXCSR(bxInstruction_c *i)
{
#if BX_CPU_LEVEL >= 6
  bx_address eaddr = BX_CPU_RESOLVE_ADDR(i);

  Bit32u new_mxcsr = read_virtual_dword(i->seg(), eaddr);
  if(new_mxcsr & ~MXCSR_MASK)
      exception(BX_GP_EXCEPTION, 0);

  BX_MXCSR_REGISTER = new_mxcsr;
#endif

  BX_NEXT_INSTR(i);
}

This is MXCSR_MASK:

#define MXCSR_MASK        (BX_CPU_THIS_PTR mxcsr_mask)

And this is how it looks like it gets initialized (init.cc):

void BX_CPU_C::reset(unsigned source)
{
// ...
    BX_CPU_THIS_PTR mxcsr_mask = 0x0000ffbf;
    if (BX_CPUID_SUPPORT_ISA_EXTENSION(BX_ISA_SSE2))
      BX_CPU_THIS_PTR mxcsr_mask |= MXCSR_DAZ;
    if (BX_CPUID_SUPPORT_ISA_EXTENSION(BX_ISA_MISALIGNED_SSE))
      BX_CPU_THIS_PTR mxcsr_mask |= MXCSR_MISALIGNED_EXCEPTION_MASK;

Could you dump the value that's getting loaded?

Cheers

[Micky-Thongam]

state of mxcsr (mem.dmp) :
kd> r mxcsr
mxcsr=00001fa1

here: ldmxcsr dword ptr [rsp+8]
I dumped rsp+8 and its
[rsp+8] == 0x000000000000BFA1

[0vercl0k]

Okay, I made a mistake; the mxcsr_mask is actually set by bxcpu when we initialize the register via cpu_set_mxcsr_mask:

BOCHSAPI void cpu_set_mxcsr_mask(unsigned id, Bit32u v) {
    BX_CPU(id)->mxcsr_mask = v;
}

So in your case, mxcsr_mask is 0b0001111110100001 and you are trying to load 0b1011111110100001 into it; but 0b1011111110100001 has two bits set (1<<13 & 1<<15) that aren't set in the mask, so you take a #GP:

mask =0b0001111110100001
value=0b1011111110100001
        ^ ^
        | |

Here's the description of those two bits:

/* 31|30|29|28|27|26|25|24|23|22|21|20|19|18|17|16
 * ==|==|=====|==|==|==|==|==|==|==|==|==|==|==|==  (reserved)
 *  0| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0|MM| 0
 *
 * 15|14|13|12|11|10| 9| 8| 7| 6| 5| 4| 3| 2| 1| 0
 * ==|==|=====|==|==|==|==|==|==|==|==|==|==|==|==
 * FZ| R C |PM|UM|OM|ZM|DM|IM|DZ|PE|UE|OE|ZE|DE|IE
 */

/* MXCSR REGISTER FIELDS DESCRIPTION */

/*
 * RC 13-14 Floating-Point Rounding Control         00
 * FZ 15    Flush-to-Zero for Masked Underflow      0
 */

To verify all of the above, can you change your regs.json so that mxcsr_mask is 0xbfa1 and not 0x1fa1, and re-run your testcase?

Cheers