Re: [Qemu-devel] [PATCH] fix fdiv instruction

[G 3]

On Jun 26, 2018, at 9:49 AM, Richard Henderson wrote:

> On 06/25/2018 03:23 PM, Programmingkid wrote:
> > > On Jun 25, 2018, at 5:08 PM, Richard Henderson  
> > > <address@hidden> wrote:
> > >
> > > On Mon, Jun 25, 2018, 08:23 G 3 <address@hidden> wrote:
> > > > Try
> > > >
> > > >     uint64_t expected_answer = 0xffff0000deadbeef;
> > > >     ...
> > > >     c.i = expected_answer;
> > > >     asm volatile("fdiv %0, %1, %2" : "+f"(c.d) : "f"(1.0),  
> > > > "f"(0.0));
> > > >
> > > > to avoid depending on uninitialized data.  (This expected value is
> > > > an SNaN with a deadbeef marker Just to be Sure.)
> > > >
> > > >
> > > > r~
> > >
> > >
> > > Ok I made this program and tried it on my iMac G5 (PowerPC 970).
> > >
> > > #include <stdio.h>
> > > #include <stdint.h>
> > > #include <inttypes.h>
> > >
> > > // Used to convert unsigned integer <--> double
> > > union Converter
> > > {
> > >      double d;
> > >      uint64_t i;
> > > };
> > > typedef union Converter Converter;
> > >
> > > int main (int argc, const char * argv[]) {
> > >      Converter answer;
> > >      answer.i = 0xffff0000deadbeef;
> > >      //asm volatile("mtfsb1 27"); /* Set ZE bit */
> > >      asm volatile("fdiv %0, %1, %2" : "=f"(answer.d) : "f"(1.0),
> > > "f"(0.0));
> > >
> > > Need +f for inout operand.
> > > This didn't test what you expected.
> >
> > What do you mean by inout operand?
>
> An operand that is both input and output.
>
> > If you could send me some sample code I will test it out.
>
> I did, you just didn't read it properly.
> Here it is again:
>
>   asm volatile("fdiv %0, %1, %2"
>                : "+f"(answer.d) : "f"(1.0), "f"(0.0));
>                  ^^^^
>
> r~


I used the +f constraint and did see different results. Computer used  
was an iMac G5 (PowerPC 970).

If FPSCR[ZE] is set:
        answer = 0xffff0000deadbeef  (changed from 0x0)

if FPSCR[ZE] is not set:
        answer = 0x7ff0000000000000  (no change from before)


I then tried this program in QEMU without any of my patches applied.  
Here is the result:

If FPSCR[ZE] is set or not set, answer = 0x7ff0000000000000. This  
indicates to me that the fdiv instruction needs a little work. This  
is what I think should happen. If division by zero takes  place and  
the FPSCR[ZE] bit is set, then the value in the destination register  
should not be altered (rather than returning zero).


Here is the program used:

#include <stdio.h>
#include <stdint.h>
#include <inttypes.h>

// Used to convert unsigned integer <--> double
union Converter
{
    double d;
    uint64_t i;
};

typedef union Converter Converter;

int main (int argc, const char * argv[]) {

    Converter answer;
    answer.i = 0xffff0000deadbeef;
    //asm volatile("mtfsb1 27"); /* Set ZE bit */
    asm volatile("fdiv %0, %1, %2" : "+f"(answer.d) : "f"(1.0),  
"f"(0.0));
    printf("answer = 0x%" PRIx64 "\n", answer.i);
    return 0;
}