On Sat, Nov 23, 2013 at 5:48 PM, Uros Bizjak <ubiz...@gmail.com> wrote:
>> Thanks for lookin at this. I am a real newcomer to 387, and it took me a
>> long time perusing the Intel doc, as well as glibc sources, to come up with
>> that. The “reference” implementation of these FPU functions, the one I am
>> confident in, is that in config/fpu-glibc.h: i.e., the functions in
>> config/fpu-i387.h should have the same effect that config/fpu-glibc.h on
>> i386/x86_64 hardware.
>>
>> I’ll reply to your comments, but in some cases I was not sure exactly what
>> you were saying… thanks for your help, and patience!
>>
>>
>>> @@ -136,16 +165,54 @@ set_fpu (void)
>>> __asm__ __volatile__ ("%vstmxcsr\t%0" : "=m" (cw_sse));
>>>
>>> /* The SSE exception masks are shifted by 7 bits. */
>>> - cw_sse |= _FPU_MASK_ALL << 7;
>>> - cw_sse &= ~(excepts << 7);
>>> -
>>> - /* Clear stalled exception flags. */
>>> - cw_sse &= ~_FPU_EX_ALL;
>>>
>>> You have to clear stalled SSE exceptions here. Their flags are in LSB
>>> bits, so their position is different than the position of exception
>>> mask bits in the control word.
>>
>> So, if I get you right, I should restore the "cw_sse &= ~_FPU_EX_ALL”, which
>> I had mistakenly removed.
>> But I’m looking at glibc-2.18/sysdeps/x86_64/fpu/feenablxcpt.c and
>> fedisblxcpt.c, and it doesn’t seem to be done there.
>
> The idea was that since control word is changed, status word should be
> cleared. But since stalled flags won't raise an exception, it actually
> doesn't matter, although it looks nicer in a debugger. However, if you
> remove SSE clear, you should also remove fnclex from x87 code.
Actually, I was wrong. You have to clear stalled flags.
Please consider following test:
--cut here--
#include <fenv.h>
int main(void)
{
unsigned short cw;
/* Raise FE_DIVBYZERO */
volatile float d = 0.0f;
volatile float r = 1.0f / d;
#if GLIBC
feenableexcept (FE_DIVBYZERO);
#else
__asm__ __volatile__ ("fstcw\t%0" : "=m" (cw));
cw |= FE_ALL_EXCEPT;
cw &= ~FE_DIVBYZERO;
__asm__ __volatile__ ("fnclex\n\tfldcw\t%0" : : "m" (cw));
#endif
/* Raise FE_INEXACT */
d = 3.0f;
r = 1.0f / d;
return 0;
}
--cut here--
The test (compiled with -m32 or -mfpmath=387 due to x87 assembly,
linked against -lm will generate erroneous exception when -DGLIBC is
added to compile flags.
So, it looks to me that glibc has a bug here. Oh, and contrary to
claims in glibc sources, the above test raises only FE_INEXACT
exception.
I have added Joseph to Cc due to glibc issues.
Uros.
